scripts: export per-voice prefix/suffix embeddings

New tool + generated artefacts so the on-device voice spinner can now
hot-swap between all 8 voices — previously only Damien's prefix/suffix
were present in the model dir, and the tablet fell back to him
regardless of selection.

scripts/export_voice_prefix_suffix.py runs Qwen3TTS's voice-clone
path under a forward hook, captures the first prefill call's 1024-dim
talker input embeddings, aborts the rest of the (very slow on CPU)
decode via a sentinel exception, and slices out the first 9 vectors
as <name>_voice_prefix.bin and the last 2 as <name>_voice_suffix.bin.
Validated against the shipped damien_voice_prefix.bin: using
damien_15s_24k.wav as the reference audio, max|diff| = 0, so the
extraction matches the original tooling bit-for-bit.

Generated and adb-pushed to
/data/local/tmp/kazeia/models/qwen3-tts-npu/:
  amir / didier / elodie / jerome / richard / sid / zelda
  (+ re-generated damien from the canonical 15s_24k reference)

Qwen3TtsEngine.setVoice (already wired) reads <voice>_voice_prefix.bin
/ <voice>_voice_suffix.bin by basename, so voice changes now take
effect from the next synthesized segment with no app restart.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

executorch-patches/llm_in_process_jni.patch

@@ -0,0 +1,72 @@
+diff --git a/backends/qualcomm/CMakeLists.txt b/backends/qualcomm/CMakeLists.txt
+index e93731e..4951e1d 100644
+--- a/backends/qualcomm/CMakeLists.txt
++++ b/backends/qualcomm/CMakeLists.txt
+@@ -308,8 +308,8 @@ if(${CMAKE_SYSTEM_PROCESSOR} MATCHES Hexagon)
+   )
+ endif()
+ 
+-# QNN pybind
+-if(${CMAKE_SYSTEM_PROCESSOR} MATCHES "x86_64")
++# QNN pybind — host Python bindings, not for Android cross-compile
++if(${CMAKE_SYSTEM_PROCESSOR} MATCHES "x86_64" AND NOT ANDROID)
+   add_subdirectory(
+     ${EXECUTORCH_SOURCE_DIR}/third-party/pybind11
+     ${CMAKE_CURRENT_BINARY_DIR}/pybind11
+diff --git a/extension/android/jni/jni_layer_llama.cpp b/extension/android/jni/jni_layer_llama.cpp
+index 45f2414..ae3d79f 100644
+--- a/extension/android/jni/jni_layer_llama.cpp
++++ b/extension/android/jni/jni_layer_llama.cpp
+@@ -171,14 +171,44 @@ class ExecuTorchLlmJni : public facebook::jni::HybridClass<ExecuTorchLlmJni> {
+           model_path->toStdString().c_str(),
+           data_files_vector,
+           executorch::extension::Module::LoadMode::MmapUseMlockIgnoreErrors);
+-      std::string decoder_model = "llama3"; // use llama3 for now
+-      runner_ = std::make_unique<example::Runner<uint16_t>>( // QNN runner
+-          std::move(module),
+-          decoder_model.c_str(),
+-          model_path->toStdString().c_str(),
+-          tokenizer_path->toStdString().c_str(),
+-          "",
+-          "");
++      std::string decoder_model = "qwen3"; // Kazeia: our .pte was exported with --decoder_model qwen3-4b
++
++      // Mirror qnn_llama_runner.cpp main(): pick the Runner<T> template based
++      // on the model's get_kv_io_bit_width metadata. The 16-bit KV I/O models
++      // were introduced after the 8-bit ones, and using the wrong T treats
++      // KV-cache bytes as the wrong width → garbage logits → gibberish output.
++      example::KvBitWidth kv_bitwidth = example::KvBitWidth::kWidth8;
++      if (module->method_names()->count("get_kv_io_bit_width") > 0) {
++        kv_bitwidth = static_cast<example::KvBitWidth>(
++            module->get("get_kv_io_bit_width").get().toScalar().to<int64_t>());
++      }
++      // Auto-detect eval_mode: kv-only (0) if the .pte only carries
++      // kv_forward, hybrid (1) if it also has prefill_forward (which lets the
++      // runner batch the prompt prefill — TTFT drops from ~52 ms/token to
++      // sub-ms after the one-shot prefill graph). Same JNI binary works with
++      // both export modes, no code change needed when the .pte is upgraded.
++      int eval_mode = 0;
++      if (module->method_names()->count("prefill_forward") > 0) {
++        eval_mode = 1; // EvalMode::kHybrid
++      }
++      auto make_runner = [&](auto sample) -> std::unique_ptr<llm::IRunner> {
++        using T = decltype(sample);
++        return std::make_unique<example::Runner<T>>(
++            std::move(module),
++            decoder_model.c_str(),
++            model_path->toStdString().c_str(),
++            tokenizer_path->toStdString().c_str(),
++            /* performance_output_path */ "",
++            /* dump_logits_path */ "",
++            /* temperature */ 0.0f, // greedy
++            eval_mode,
++            /* shared_buffer */ true);
++      };
++      if (kv_bitwidth == example::KvBitWidth::kWidth16) {
++        runner_ = make_runner(uint16_t{0});
++      } else {
++        runner_ = make_runner(uint8_t{0});
++      }
+       model_type_category_ = MODEL_TYPE_CATEGORY_LLM;
+ #endif
+ #if defined(EXECUTORCH_BUILD_MEDIATEK)

executorch-patches/qwen3_4b_decoder.patch

@@ -1,5 +1,5 @@
 diff --git a/examples/qualcomm/oss_scripts/llama/__init__.py b/examples/qualcomm/oss_scripts/llama/__init__.py
-index 963db6e..953dc4c 100644
+index 963db6e..9ccfdd0 100644
 --- a/examples/qualcomm/oss_scripts/llama/__init__.py
 +++ b/examples/qualcomm/oss_scripts/llama/__init__.py
 @@ -25,9 +25,14 @@ from executorch.examples.models.granite import (
@@ -20,7 +20,7 @@ index 963db6e..953dc4c 100644
  from executorch.examples.models.qwen2_5 import (
      convert_weights as convert_qwen2_5_weights,
  )
-@@ -479,6 +484,34 @@ class Qwen3_1_7B(LLMModelConfig):
+@@ -479,6 +484,37 @@ class Qwen3_1_7B(LLMModelConfig):
      quant_recipe = Qwen3_1_7BQuantRecipe
  
  
@@ -40,10 +40,13 @@ index 963db6e..953dc4c 100644
 +    convert_weights = convert_qwen3_weights
 +    transform_weight = False
 +    instruct_model = True
-+    # Bumped to 2 to halve peak host RAM during QNN compile (4B at sharding=1
++    # num_sharding=1 for hybrid mode: sharding=2 produces a multi-context
-+    # OOMed on a 62 GB box, peak anon-rss 46 GB). At sharding=2 each shard
++    # .pte (2 graphs × 2 shards = 4 contexts) that the LlmModule load path
-+    # compile fits comfortably; runner stitches them at load time.
++    # can't restore (error 5010 "Context group 1 does not exist"). With
-+    num_sharding = 2
++    # sharding=1 the hybrid export needs ~46 GB RAM peak — the 192 GB swap
++    # on /swapfile handles this; compile takes ~80 min wall but completes
++    # cleanly. Single-context .pte loads fine through the JNI runner.
++    num_sharding = 1
 +    masked_softmax = True
 +    seq_mse_candidates = 0
 +    r1 = False

kazeia-android/app/src/main/AndroidManifest.xml

@@ -5,6 +5,7 @@
     <uses-permission android:name="android.permission.RECORD_AUDIO" />
     <uses-permission android:name="android.permission.FOREGROUND_SERVICE" />
     <uses-permission android:name="android.permission.FOREGROUND_SERVICE_MICROPHONE" />
+    <uses-permission android:name="android.permission.FOREGROUND_SERVICE_MEDIA_PLAYBACK" />
     <uses-permission android:name="android.permission.FOREGROUND_SERVICE_SPECIAL_USE" />
     <uses-permission android:name="android.permission.POST_NOTIFICATIONS" />
     <uses-permission android:name="android.permission.READ_EXTERNAL_STORAGE" />
@@ -50,7 +51,7 @@
 
         <service
             android:name=".service.KazeiaService"
-            android:foregroundServiceType="microphone|specialUse"
+            android:foregroundServiceType="microphone|mediaPlayback|specialUse"
             android:exported="true">
             <property
                 android:name="android.app.PROPERTY_SPECIAL_USE_FGS_SUBTYPE"

kazeia-android/app/src/main/java/com/kazeia/llm/ExecuTorchLlmEngine.kt

@@ -1,43 +1,49 @@
 package com.kazeia.llm
 
+import android.content.Context
 import android.util.Log
 import com.kazeia.core.*
 import kotlinx.coroutines.Dispatchers
 import kotlinx.coroutines.withContext
 import java.io.File
+import org.pytorch.executorch.extension.llm.LlmCallback
+import org.pytorch.executorch.extension.llm.LlmModule
 
 /**
- * LLM Engine using ExecuTorch + QNN backend via subprocess.
+ * LLM Engine using ExecuTorch LlmModule in-process — **no root required**.
- * Calls qnn_llama_runner binary with root access (Magisk su).
+ *
+ * Runs Qwen3-4B via `org.pytorch.executorch.extension.llm.LlmModule`, which
+ * wraps the same C++ TextLlmRunner as the standalone qnn_llama_runner binary
+ * but inside the app's own process. The QNN HTP backend works because the
+ * DSP fastrpc service accepts the Zygote-forked app process (unlike
+ * ProcessBuilder-spawned subprocesses which lose supplementary GIDs on exec
+ * and get rejected by the fastrpc credential checks).
+ *
+ * Model + tokenizer live in /data/local/tmp/kazeia-et/ (readable by the app
+ * on this device's permissive SELinux policy). libexecutorch.so + QNN libs
+ * are bundled in jniLibs.
  *
  * Current tablet config: Qwen3-4B KV-mode, ~18-22 tok/s on Hexagon V79
  * (Snapdragon 8 Elite), TTFT 0.9 s, RSS 1.76 GB.
- *
- * Why root: the runner binary plus its QNN v2.42 .so deps live in
- * /data/local/tmp/kazeia-et/ (shell_data_file SELinux context). Untrusted
- * apps can't exec binaries from there. The Hexagon DSP fastrpc service also
- * refuses to load the v2.42 Skel from the app's own files dir — only from
- * nativeLibraryDir — but that dir already holds the TTS stack's v2.31 Skel
- * (same filename, different version, can't coexist). Rebuilding everything
- * against one QNN version would eliminate the conflict, but would require
- * re-exporting the TTS .pte with the new runtime (tooling currently broken
- * on the flatc schema/dataclass mismatch in the qnn_venv).
  */
 class ExecuTorchLlmEngine(
+    private val context: Context,
     private val onLog: ((String) -> Unit)? = null
 ) : LlmEngine {
 
     companion object {
         private const val TAG = "ExecuTorchLLM"
-        private const val RUNNER_DIR = "/data/local/tmp/kazeia-et"
+        // /no_think disables Qwen3's chain-of-thought block. Compact wording
-        // /no_think disables Qwen3's chain-of-thought block so the full token
+        // keeps prefill cost low: this prompt is ~25 tokens vs ~55 in the
-        // budget goes to the actual answer (without it, 120-200 tokens get
+        // earlier verbose version → saves ~1.5 s of TTFT in kv-only mode.
-        // consumed by <think>…</think> leaving nothing to speak).
+        private const val SYSTEM_PROMPT = "Tu es Kazeia, à l'écoute en français. Réponds en 1-2 phrases courtes, sans raisonnement. /no_think"
-        // Short-response directive keeps TTS latency manageable — each sentence
+
-        // costs ~3-5 s on the .pte path, so 1-2 sentences is the sweet spot.
+        private const val MODEL_DIR = "/data/local/tmp/kazeia-et"
-        private const val SYSTEM_PROMPT = "Tu es Kazeia, un compagnon bienveillant d'écoute émotionnelle. Réponds toujours en français, en 1 ou 2 phrases courtes (40 mots maximum). Pas de raisonnement, donne directement la réponse. /no_think"
+        private const val MODEL_PATH = "$MODEL_DIR/hybrid_llama_qnn.pte"
+        private const val TOKENIZER_PATH = "$MODEL_DIR/tokenizer.json"
     }
 
+    private var llmModule: LlmModule? = null
     private var modelName = ""
     private var loaded = false
 
@@ -48,77 +54,152 @@ class ExecuTorchLlmEngine(
 
     override suspend fun load(modelPath: String, config: LlmConfig) {
         withContext(Dispatchers.IO) {
-            val check = execRoot("ls $RUNNER_DIR/qnn_llama_runner $RUNNER_DIR/hybrid_llama_qnn.pte $RUNNER_DIR/tokenizer.json 2>&1")
+            if (!File(MODEL_PATH).exists()) {
-            if (check.contains("No such file")) {
+                nlog("ERROR: model not found at $MODEL_PATH")
-                nlog("ERROR: runner or model not found in $RUNNER_DIR")
+                return@withContext
+            }
+            if (!File(TOKENIZER_PATH).exists()) {
+                nlog("ERROR: tokenizer not found at $TOKENIZER_PATH")
                 return@withContext
             }
 
-            deployRunnerScript()
+            try {
+                val t0 = System.currentTimeMillis()
+                // MODEL_TYPE_QNN_LLAMA=4 selects the Qualcomm runner path in
+                // jni_layer_llama.cpp, which uses example::Runner (same code
+                // as the qnn_llama_runner binary) instead of the generic
+                // TextLLMRunner. Our .pte was exported with
+                // --decoder_model qwen3-4b which requires this path.
+                val MODEL_TYPE_QNN_LLAMA = 4
+                llmModule = LlmModule(MODEL_TYPE_QNN_LLAMA, MODEL_PATH, TOKENIZER_PATH, 0.7f)
+                nlog("LlmModule instantiated in ${System.currentTimeMillis() - t0}ms")
 
-            writeFileRoot("$RUNNER_DIR/outputs/prompt.b64",
+                // Load the PTE into QNN HTP (calls the native load()).
-                android.util.Base64.encodeToString("Bonjour".toByteArray(), android.util.Base64.NO_WRAP))
+                val loadResult = llmModule!!.load()
-            if (SYSTEM_PROMPT.isNotEmpty()) {
+                if (loadResult != 0) {
-                writeFileRoot("$RUNNER_DIR/outputs/system.b64",
+                    nlog("ERROR: LlmModule.load() returned $loadResult")
-                    android.util.Base64.encodeToString(SYSTEM_PROMPT.toByteArray(), android.util.Base64.NO_WRAP))
+                    llmModule = null
-            } else {
+                    return@withContext
-                execRoot("rm -f $RUNNER_DIR/outputs/system.b64")
                 }
-            val test = execRoot("sh $RUNNER_DIR/run_llm.sh 0.0 80 2>&1")
+                nlog("LlmModule loaded in ${System.currentTimeMillis() - t0}ms total")
 
-            if (test.contains("Generated Tokens") || test.contains("Rate:")) {
                 loaded = true
-                val rateMatch = Regex("Generated \\d+ tokens:.*Rate:\\s+([\\d.]+)").find(test)
+                modelName = "Qwen3-4B LlmModule"
-                val rate = rateMatch?.groupValues?.get(1) ?: "?"
-                modelName = "Qwen3 (${rate} tok/s NPU)"
                 nlog("Ready: $modelName")
-            } else {
+            } catch (e: Throwable) {
-                nlog("ERROR: test failed: ${test.takeLast(200)}")
+                nlog("ERROR: LlmModule init failed: ${e.javaClass.simpleName}: ${e.message}")
+                llmModule = null
             }
         }
     }
 
-    override fun isLoaded(): Boolean = loaded
+    override fun isLoaded(): Boolean = loaded && llmModule != null
 
     override suspend fun generate(
         prompt: String,
         params: SamplingParams,
         onToken: ((String) -> Boolean)?
     ): GenerationResult = withContext(Dispatchers.IO) {
-        if (!loaded) throw IllegalStateException("Model not loaded")
+        val mod = llmModule ?: throw IllegalStateException("Model not loaded")
 
         val startTime = System.currentTimeMillis()
-
+        val fullPrompt = buildChatTemplate(prompt)
-        writeFileRoot("$RUNNER_DIR/outputs/prompt.b64",
-            android.util.Base64.encodeToString(prompt.toByteArray(), android.util.Base64.NO_WRAP))
-        if (SYSTEM_PROMPT.isNotEmpty()) {
-            writeFileRoot("$RUNNER_DIR/outputs/system.b64",
-                android.util.Base64.encodeToString(SYSTEM_PROMPT.toByteArray(), android.util.Base64.NO_WRAP))
-        } else {
-            execRoot("rm -f $RUNNER_DIR/outputs/system.b64")
-        }
-
         nlog("Prompt: '${prompt.take(80)}'")
 
+        val responseBuilder = StringBuilder()
+        var firstTokenMs = -1L
+        // Track whether we're inside a <think>…</think> block so the upstream
+        // SentenceStreamer / TTS doesn't get fed reasoning tokens. Even with
+        // /no_think in the system prompt Qwen3 still emits empty <think></think>
+        // wrappers for ~3 tokens before the real answer.
+        var inThink = false
+        val tokenScan = StringBuilder()  // small lookahead to spot tag boundaries
+
+        // Singleton special tokens that should never reach the TTS streamer
+        // (they leak when the model wraps its reply or signals end-of-turn).
+        val stripTokens = listOf("<|im_start|>", "<|im_end|>", "<|endoftext|>")
+        val maxTagLen = listOf("<think>", "</think>", "<|im_start|>", "<|im_end|>", "<|endoftext|>")
+            .maxOf { it.length }
+
+        val cb = object : LlmCallback {
+            override fun onResult(result: String) {
+                if (firstTokenMs < 0) firstTokenMs = System.currentTimeMillis() - startTime
+                responseBuilder.append(result)
+
+                // Forward to caller only outside <think> blocks, and strip
+                // singleton special tokens. We accumulate a tiny lookahead buffer
+                // so tag tokens that arrive split ("<thi", "nk>") still match.
+                tokenScan.append(result)
+                while (true) {
+                    if (!inThink) {
+                        val open = tokenScan.indexOf("<think>")
+                        if (open < 0) {
+                            // No <think> open pending — strip any singleton tokens
+                            // that fully landed in the buffer, then flush prose
+                            // up to a safe point preserving lookahead.
+                            for (tok in stripTokens) {
+                                var idx = tokenScan.indexOf(tok)
+                                while (idx >= 0) {
+                                    tokenScan.delete(idx, idx + tok.length)
+                                    idx = tokenScan.indexOf(tok)
+                                }
+                            }
+                            val safe = tokenScan.length - maxTagLen
+                            if (safe > 0) {
+                                onToken?.invoke(tokenScan.substring(0, safe))
+                                tokenScan.delete(0, safe)
+                            }
+                            break
+                        }
+                        // Flush the prose before the <think> tag, then enter think mode.
+                        if (open > 0) onToken?.invoke(tokenScan.substring(0, open))
+                        tokenScan.delete(0, open + "<think>".length)
+                        inThink = true
+                    } else {
+                        val close = tokenScan.indexOf("</think>")
+                        if (close < 0) {
+                            // Drop all buffered chars except a small tail in case
+                            // the closing tag is split across tokens.
+                            val keep = "</think>".length - 1
+                            if (tokenScan.length > keep) tokenScan.delete(0, tokenScan.length - keep)
+                            break
+                        }
+                        tokenScan.delete(0, close + "</think>".length)
+                        inThink = false
+                    }
+                }
+            }
+            override fun onStats(stats: String) {
+                nlog("stats: ${stats.take(200)}")
+            }
+        }
+
         val seqLen = minOf(params.maxNewTokens, 512)
-        val output = execRoot("sh $RUNNER_DIR/run_llm.sh ${params.temperature} $seqLen 2>&1")
+        val rc = try {
+            // echo=false so onResult() only receives the generated completion,
+            // not the prompt tokens echoed back — otherwise the sentence
+            // streamer would feed '<|im_start|>user …' to the TTS.
+            mod.generate(fullPrompt, seqLen, cb, /* echo */ false)
+        } catch (e: Throwable) {
+            nlog("generate() threw: ${e.message}")
+            -1
+        }
 
-        val tokenCount = Regex("Generated Tokens:\\s+(\\d+)").find(output)
+        // Drain any leftover prose buffered during <think>-suppression so the
-            ?.groupValues?.get(1)?.toIntOrNull() ?: 0
+        // last sentence reaches the TTS even if it ran past the closing tag.
-        val rate = Regex("Generated \\d+ tokens:.*Rate:\\s+([\\d.]+)").find(output)
+        if (!inThink && tokenScan.isNotEmpty()) {
-            ?.groupValues?.get(1)?.toFloatOrNull() ?: 0f
+            onToken?.invoke(tokenScan.toString())
-        val ttft = Regex("Time to first generated token:\\s+([\\d.]+)").find(output)
+            tokenScan.clear()
-            ?.groupValues?.get(1)?.toFloatOrNull() ?: 0f
+        }
-
-        val responseRaw = execRoot("cat $RUNNER_DIR/outputs/response.txt 2>/dev/null")
-        nlog("RAW: ${responseRaw.take(300)}")
-        val responseText = extractResponse(responseRaw)
 
         val elapsed = System.currentTimeMillis() - startTime
-        nlog("Response: '$responseText'")
+        val rawText = responseBuilder.toString()
-        nlog("Stats: ${tokenCount}tok ${rate}tok/s TTFT=${ttft}s ${elapsed}ms")
+        val responseText = cleanResponse(rawText)
+        val tokenCount = rawText.length / 4  // rough estimate without a tokenizer
+        val rate = if (elapsed > 0) (tokenCount * 1000f) / elapsed else 0f
 
-        onToken?.invoke(responseText)
+        nlog("Response: '${responseText.take(80)}'")
+        nlog("Stats: rc=$rc ~${tokenCount}tok ~${"%.1f".format(rate)}tok/s TTFT=${firstTokenMs}ms total=${elapsed}ms")
 
         GenerationResult(
             text = responseText,
@@ -128,20 +209,32 @@ class ExecuTorchLlmEngine(
         )
     }
 
-    private fun extractResponse(raw: String): String {
+    /**
+     * Qwen3 chat template matching qnn_llama_runner.cpp's get_formatted_prompt()
+     * for DecoderModelVersion::kQwen3. Note the user-first-then-system ordering
+     * (quirky but required — the runner binary produces the same layout and our
+     * .pte was trained with it). Terminates with `<|im_start|>assistant` with
+     * no trailing newline, matching the binary exactly.
+     */
+    private fun buildChatTemplate(userInput: String): String {
+        val sb = StringBuilder()
+        sb.append("<|im_start|>user\n").append(userInput).append("<|im_end|>\n")
+        if (SYSTEM_PROMPT.isNotEmpty()) {
+            sb.append("<|im_start|>system\n").append(SYSTEM_PROMPT).append("<|im_end|>\n")
+        }
+        sb.append("<|im_start|>assistant")
+        return sb.toString()
+    }
+
+    /** Strip <think>…</think>, special tokens, and leading/trailing whitespace. */
+    private fun cleanResponse(raw: String): String {
         var text = raw
         val thinkEnd = text.indexOf("</think>")
         if (thinkEnd >= 0) {
             text = text.substring(thinkEnd + "</think>".length)
-        } else {
+        } else if (text.indexOf("<think>") >= 0) {
-            val thinkStart = text.indexOf("<think>")
+            nlog("WARN: <think> block never closed")
-            val assistantTag = text.indexOf("assistant")
-            if (thinkStart >= 0) {
-                nlog("WARN: <think> block never closed, no response generated")
             return ""
-            } else if (assistantTag >= 0) {
-                text = text.substring(assistantTag + "assistant".length)
-            }
         }
         return text
             .replace("<|im_start|>", "")
@@ -152,82 +245,9 @@ class ExecuTorchLlmEngine(
             .trim()
     }
 
-    private fun deployRunnerScript() {
-        val script = """
-#!/bin/sh
-cd $RUNNER_DIR
-export LD_LIBRARY_PATH=$RUNNER_DIR
-export ADSP_LIBRARY_PATH=$RUNNER_DIR
-
-TEMP=${'$'}1
-SEQ_LEN=${'$'}2
-
-PROMPT=${'$'}(base64 -d $RUNNER_DIR/outputs/prompt.b64)
-
-rm -f $RUNNER_DIR/outputs/response.txt
-
-SYSTEM_ARGS=""
-if [ -s $RUNNER_DIR/outputs/system.b64 ]; then
-  SYSTEM=${'$'}(base64 -d $RUNNER_DIR/outputs/system.b64)
-  SYSTEM_ARGS="--system_prompt"
-fi
-
-if [ -n "${'$'}SYSTEM_ARGS" ]; then
-  exec ./qnn_llama_runner \
-    --model_path hybrid_llama_qnn.pte \
-    --tokenizer_path tokenizer.json \
-    --decoder_model_version qwen3 \
-    --output_path $RUNNER_DIR/outputs/response.txt \
-    --performance_output_path $RUNNER_DIR/outputs/perf.txt \
-    --shared_buffer \
-    --system_prompt "${'$'}SYSTEM" \
-    --prompt "${'$'}PROMPT" \
-    --temperature ${'$'}TEMP \
-    --seq_len ${'$'}SEQ_LEN \
-    --eval_mode 0
-else
-  exec ./qnn_llama_runner \
-    --model_path hybrid_llama_qnn.pte \
-    --tokenizer_path tokenizer.json \
-    --decoder_model_version qwen3 \
-    --output_path $RUNNER_DIR/outputs/response.txt \
-    --performance_output_path $RUNNER_DIR/outputs/perf.txt \
-    --shared_buffer \
-    --prompt "${'$'}PROMPT" \
-    --temperature ${'$'}TEMP \
-    --seq_len ${'$'}SEQ_LEN \
-    --eval_mode 0
-fi
-""".trimIndent()
-
-        writeFileRoot("$RUNNER_DIR/run_llm.sh", script)
-        execRoot("chmod 755 $RUNNER_DIR/run_llm.sh")
-    }
-
     override fun release() {
+        try { llmModule?.resetNative() } catch (_: Throwable) {}
+        llmModule = null
         loaded = false
     }
-
-    private fun writeFileRoot(path: String, content: String) {
-        try {
-            val process = Runtime.getRuntime().exec(arrayOf("su", "-c", "cat > $path"))
-            process.outputStream.bufferedWriter().use { it.write(content) }
-            process.waitFor()
-        } catch (e: Exception) {
-            Log.e(TAG, "writeFileRoot failed: ${e.message}")
-        }
-    }
-
-    private fun execRoot(cmd: String): String {
-        return try {
-            val process = Runtime.getRuntime().exec(arrayOf("su", "-c", cmd))
-            val result = process.inputStream.bufferedReader().readText()
-            val error = process.errorStream.bufferedReader().readText()
-            process.waitFor()
-            if (error.isNotEmpty() && result.isEmpty()) error else result
-        } catch (e: Exception) {
-            Log.e(TAG, "execRoot failed: ${e.message}")
-            ""
-        }
-    }
 }

kazeia-android/app/src/main/java/com/kazeia/service/KazeiaPipeline.kt

@@ -142,14 +142,36 @@ class KazeiaPipeline {
      * the echo-mode playback through the same path — otherwise each TTS
      * site reimplemented the "streaming-or-fallback" dispatch.
      */
-    suspend fun speakText(text: String) {
+    suspend fun speakText(
+        text: String,
+        // Fires the instant each synthesized sentence starts playing
+        // through the speaker, with the sentence text, audio duration,
+        // and a per-ENVELOPE_WINDOW_MS RMS envelope. Used by
+        // processLlmResponse to defer the KAZEIA chat bubble appearance
+        // until sound is audible, pace word-by-word reveal inside the
+        // bubble, and drive the AudioVisualizerView orb.
+        onSegmentPlaying: ((
+            sentence: String,
+            durationMs: Long,
+            rmsEnvelope: FloatArray,
+            spectrogram: Array<FloatArray>
+        ) -> Unit)? = null
+    ) {
         val ttsEngine = tts ?: return
         _pipelineState.value = PipelineState.Speaking
         try {
             val qwen = ttsEngine as? com.kazeia.tts.Qwen3TtsEngine
             if (qwen != null) {
+                qwen.onSegmentPlaying = onSegmentPlaying
                 qwen.startStreamingSession()
-                val streamer = com.kazeia.tts.SentenceStreamer { s -> qwen.enqueueSentence(s) }
+                val streamer = com.kazeia.tts.SentenceStreamer { raw ->
+                    // Strip emoji / non-speakable pictographs before TTS
+                    // so a standalone "😊" doesn't become its own noisy
+                    // segment. The chat bubble keeps the original text —
+                    // only the audio path sees the cleaned version.
+                    val spoken = stripNonSpeakable(raw).trim()
+                    if (spoken.isNotEmpty()) qwen.enqueueSentence(spoken)
+                }
                 streamer.append(text)
                 streamer.flush()
                 qwen.endStreamingSession()
@@ -168,6 +190,41 @@ class KazeiaPipeline {
         _messages.value = _messages.value + msg
     }
 
+    /**
+     * Drop emoji + dingbat + pictographic characters so the TTS engine
+     * doesn't try to synthesize them. Covers the main Unicode emoji
+     * blocks (Miscellaneous Symbols, Dingbats, Emoticons, Transport,
+     * Supplemental Symbols and Pictographs, etc.) plus variation
+     * selectors and zero-width joiners that tag emoji sequences.
+     * Keeps everything in the Basic Latin / Latin-1 / Latin Extended
+     * ranges + common French punctuation untouched.
+     */
+    private fun stripNonSpeakable(text: String): String {
+        val sb = StringBuilder(text.length)
+        var i = 0
+        while (i < text.length) {
+            val cp = text.codePointAt(i)
+            val skip = when {
+                cp in 0x2600..0x27BF -> true              // misc symbols + dingbats
+                cp in 0x1F300..0x1F5FF -> true            // pictographs
+                cp in 0x1F600..0x1F64F -> true            // emoticons
+                cp in 0x1F680..0x1F6FF -> true            // transport
+                cp in 0x1F700..0x1F77F -> true            // alchemical
+                cp in 0x1F780..0x1F7FF -> true            // geometric extended
+                cp in 0x1F800..0x1F8FF -> true            // supplemental arrows-c
+                cp in 0x1F900..0x1F9FF -> true            // supplemental pictographs
+                cp in 0x1FA00..0x1FAFF -> true            // symbols & pictographs extended-A
+                cp == 0x200D -> true                       // zero-width joiner
+                cp in 0xFE00..0xFE0F -> true              // variation selectors
+                cp in 0x1F1E6..0x1F1FF -> true            // regional indicators (flags)
+                else -> false
+            }
+            if (!skip) sb.appendCodePoint(cp)
+            i += Character.charCount(cp)
+        }
+        return sb.toString()
+    }
+
     fun log(msg: String) {
         Log.i(TAG, msg)
         val time = java.text.SimpleDateFormat("HH:mm:ss.SSS", java.util.Locale.FRANCE)

kazeia-android/app/src/main/java/com/kazeia/service/KazeiaService.kt

@@ -83,6 +83,34 @@ class KazeiaService : Service() {
     private val _isListening = MutableStateFlow(false)
     val isListening: StateFlow<Boolean> = _isListening
 
+    // Drives the AudioVisualizerView orb. Pushed from the VAD loop
+    // during mic capture (mic RMS, normalized) and from the TTS engine's
+    // onSegmentPlaying callback (TTS RMS envelope per-segment). The view
+    // reads this via collectLatest in ChatActivity; the signals carry
+    // their own state so the visualizer knows whether it's idle, tracking
+    // the mic, or rendering a TTS segment.
+    sealed class VisualizerSignal {
+        object Idle : VisualizerSignal()
+        data class Listening(val micRms: Float) : VisualizerSignal()
+        data class Speaking(
+            val rmsEnvelope: FloatArray,
+            val spectrogram: Array<FloatArray>,
+            val durationMs: Long
+        ) : VisualizerSignal()
+    }
+    private val _visualizerSignal = MutableStateFlow<VisualizerSignal>(VisualizerSignal.Idle)
+    val visualizerSignal: StateFlow<VisualizerSignal> = _visualizerSignal
+
+    // Kazeia's orb color is bound to the selected voice so the user
+    // visually associates a palette with the speaker they picked. UI
+    // sets this whenever the voice spinner changes; the orb view
+    // listens via the StateFlow and tweens the current → target color.
+    private val _voiceColor = MutableStateFlow(0xFFBCA4E8.toInt())   // lavender = Damien default
+    val voiceColor: StateFlow<Int> = _voiceColor
+
+    /** Called by the UI whenever the voice selector changes. */
+    fun setVoiceColor(color: Int) { _voiceColor.value = color }
+
     private val _debugMode = MutableStateFlow(false)
     val debugMode: StateFlow<Boolean> = _debugMode
 
@@ -174,6 +202,12 @@ class KazeiaService : Service() {
                     if (!::llm.isInitialized || !llm.isLoaded()) {
                         log("Stream LLM: LLM not ready"); return@launch
                     }
+                    // Set pipeline state to Speaking so the continuous-
+                    // listening mic loop (line ~824) drops frames during
+                    // TTS playback. Without this, the mic picks up the
+                    // tablet speaker and feeds our own TTS back into STT,
+                    // creating an infinite loop.
+                    _pipelineState.value = PipelineState.Speaking
                     qwenTts.startStreamingSession()
                     val tStart = System.currentTimeMillis()
                     var firstSentenceLogged = false
@@ -199,6 +233,9 @@ class KazeiaService : Service() {
                 } catch (e: Exception) {
                     log("Stream LLM error: ${e.message}")
                     e.printStackTrace()
+                } finally {
+                    // Back to Idle so the next mic frame is accepted.
+                    _pipelineState.value = PipelineState.Idle
                 }
             }
         }
@@ -414,9 +451,17 @@ class KazeiaService : Service() {
             this, Manifest.permission.RECORD_AUDIO
         ) == PackageManager.PERMISSION_GRANTED
 
+        // FOREGROUND_SERVICE_TYPE_MEDIA_PLAYBACK is required so ColorOS (and
+        // stock Android 14+ policies) don't mute the TTS AudioTrack with
+        // "clientVolume" at ~600 ms after play(). Without it the FGS was
+        // classified as mic-only or special-use and background-audio
+        // hardening silenced it. Combine with MICROPHONE so mic input keeps
+        // working during STT.
         val fgsType = if (hasMicPermission) {
-            ServiceInfo.FOREGROUND_SERVICE_TYPE_MICROPHONE
+            ServiceInfo.FOREGROUND_SERVICE_TYPE_MICROPHONE or
+                ServiceInfo.FOREGROUND_SERVICE_TYPE_MEDIA_PLAYBACK
         } else {
+            ServiceInfo.FOREGROUND_SERVICE_TYPE_MEDIA_PLAYBACK or
                 ServiceInfo.FOREGROUND_SERVICE_TYPE_SPECIAL_USE
         }
 
@@ -450,7 +495,7 @@ class KazeiaService : Service() {
                 // TTS: try Qwen3-TTS (NPU Hexagon), fallback to Android TTS
                 _loadingState.value = LoadingState(15, "TTS Qwen3…")
                 try {
-                    val qwenTts = com.kazeia.tts.Qwen3TtsEngine(nativeLibDir) { msg -> log("[TTS] $msg") }
+                    val qwenTts = com.kazeia.tts.Qwen3TtsEngine(nativeLibDir, this@KazeiaService) { msg -> log("[TTS] $msg") }
                     qwenTts.load("$modelsDir/qwen3-tts-npu")
                     if (qwenTts.isLoaded()) {
                         tts = qwenTts
@@ -518,7 +563,7 @@ class KazeiaService : Service() {
 
                 // LLM: Qwen3-4B on Hexagon V79 via qnn_llama_runner.
                 _loadingState.value = LoadingState(50, "LLM Qwen3-4B NPU…")
-                llm = ExecuTorchLlmEngine { msg -> log(msg) }
+                llm = ExecuTorchLlmEngine(this@KazeiaService) { msg -> log(msg) }
                 try {
                     llm.load("${KazeiaApplication.MODELS_DIR}/qwen3-4b", com.kazeia.core.LlmConfig())
                 } catch (e: Exception) {
@@ -583,6 +628,16 @@ class KazeiaService : Service() {
         if (chatterbox != null) {
             chatterbox.setVoice(voicePath)
             log("Voice set to: $voicePath")
+            return
+        }
+        val qwen = tts as? com.kazeia.tts.Qwen3TtsEngine
+        if (qwen != null) {
+            // Hot-swap prefix/suffix embeddings — no model reload. Takes
+            // effect from the NEXT synthesized segment (current in-flight
+            // one, if any, finishes with the old voice since the arrays
+            // are already in its closure).
+            qwen.setVoice(voicePath)
+            log("Voice set to: $voicePath")
         }
     }
 
@@ -835,6 +890,14 @@ class KazeiaService : Service() {
                     for (s in frame) sumSq += s.toLong() * s.toLong()
                     val rms = Math.sqrt(sumSq.toDouble() / frameSize).toInt()
 
+                    // Drive the visualizer orb. Normalize with the same
+                    // sqrt squashing used for TTS so loud peaks don't
+                    // saturate and quiet speech is still visible. The
+                    // visualizer stays in Listening mode; it will swap
+                    // to Speaking or Idle when pipelineState moves on.
+                    val rmsNorm = kotlin.math.sqrt((rms / 6000f).coerceIn(0f, 1f))
+                    _visualizerSignal.value = VisualizerSignal.Listening(rmsNorm)
+
                     // Log RMS every second for calibration
                     if (frameCount % 10 == 0) {
                         Log.d(TAG, "VAD RMS=$rms (threshold=$silenceThreshold)")
@@ -1184,12 +1247,99 @@ class KazeiaService : Service() {
             log("LLM stats: ${result.tokenCount} tokens in ${result.timeMs}ms (${result.tokensPerSecond} tok/s)")
 
             if (responseText.isNotEmpty()) {
-                addMessage(ChatMessage(role = ChatMessage.Role.KAZEIA, text = responseText))
+                // Mark the pipeline as Speaking for the duration of TTS so
-                pipeline.speakText(responseText)
+                // the continuous-listening mic loop drops frames and we
+                // don't feed our own speaker output back into STT.
+                _pipelineState.value = PipelineState.Speaking
+                // Create a KAZEIA bubble up-front. Until the first TTS
+                // segment actually starts playing the bubble shows an
+                // animated "." → ".." → "..." typing indicator so the
+                // user knows Kazeia is thinking/synthesising; once the
+                // first segment plays the dots are cleared and the
+                // per-sentence word reveal takes over.
+                val bubble = ChatMessage(role = ChatMessage.Role.KAZEIA, text = ".")
+                addMessage(bubble)
+                val revealScope = kotlinx.coroutines.CoroutineScope(kotlinx.coroutines.Dispatchers.Default)
+                var revealedSoFar = ""
+                val revealJobs = mutableListOf<kotlinx.coroutines.Job>()
+                val firstSegmentSeen = java.util.concurrent.atomic.AtomicBoolean(false)
+                val typingJob = revealScope.launch {
+                    var tick = 0
+                    while (!firstSegmentSeen.get()) {
+                        val dots = ".".repeat(1 + (tick % 3))   // . → .. → ...
+                        updateMessageText(bubble.id, dots)
+                        tick++
+                        kotlinx.coroutines.delay(400)
                     }
-
+                }
+                try {
+                    pipeline.speakText(responseText) { sentence, durationMs, envelope, spectrogram ->
+                        // First segment: stop the typing indicator and
+                        // reset the bubble to empty so the word reveal
+                        // doesn't collide with the dots.
+                        if (firstSegmentSeen.compareAndSet(false, true)) {
+                            try { typingJob.cancel() } catch (_: Exception) {}
+                            updateMessageText(bubble.id, "")
+                        }
+                        // Push the envelope + spectrogram to the
+                        // visualizer at the same moment the MediaPlayer
+                        // starts playing so the orb reacts to this
+                        // segment's actual energy and the in-sphere
+                        // spectrum bars match the audio content.
+                        _visualizerSignal.value =
+                            VisualizerSignal.Speaking(envelope, spectrogram, durationMs)
+                        // Start a coroutine that appends one word at a time
+                        // over the segment's audio duration. Words are
+                        // separated on whitespace; punctuation rides with
+                        // the trailing word. The prefix (= text already
+                        // revealed from previous sentences) carries over so
+                        // earlier sentences stay on screen.
+                        val prefix = revealedSoFar
+                        val words = sentence.split(Regex("\\s+")).filter { it.isNotBlank() }
+                        revealedSoFar =
+                            if (prefix.isEmpty()) sentence
+                            else "$prefix $sentence"
+                        if (words.isEmpty()) return@speakText
+                        val perWordMs = (durationMs / words.size).coerceAtLeast(40L)
+                        val job = revealScope.launch {
+                            val sb = StringBuilder(prefix)
+                            if (prefix.isNotEmpty()) sb.append(' ')
+                            // Immediately reveal the first word so there's
+                            // no visible gap between audio start and text.
+                            sb.append(words[0])
+                            updateMessageText(bubble.id, sb.toString())
+                            for (i in 1 until words.size) {
+                                kotlinx.coroutines.delay(perWordMs)
+                                sb.append(' ').append(words[i])
+                                updateMessageText(bubble.id, sb.toString())
+                            }
+                        }
+                        revealJobs.add(job)
+                    }
+                    // After all segments finished playing, ensure the full
+                    // text is visible even if a reveal job was racing.
+                    revealJobs.forEach { try { it.join() } catch (_: Exception) {} }
+                    updateMessageText(bubble.id, responseText)
+                } finally {
+                    // Defensive: cancel the typing dots in case no
+                    // segment ever fired (e.g. the response was entirely
+                    // emojis and got stripped empty).
+                    firstSegmentSeen.set(true)
+                    try { typingJob.cancel() } catch (_: Exception) {}
                     _pipelineState.value = if (_isListening.value)
                         PipelineState.Listening else PipelineState.Idle
+                    // If we're going back to mic listening, the VAD loop
+                    // will keep pushing Listening signals; otherwise drop
+                    // to Idle so the orb settles back to its breathing
+                    // baseline.
+                    if (!_isListening.value) {
+                        _visualizerSignal.value = VisualizerSignal.Idle
+                    }
+                }
+            } else {
+                _pipelineState.value = if (_isListening.value)
+                    PipelineState.Listening else PipelineState.Idle
+            }
 
         } catch (e: Exception) {
             _aiWorkload.value = _aiWorkload.value.copy(llmActive = false)
@@ -1207,6 +1357,19 @@ class KazeiaService : Service() {
         _messages.value = _messages.value + message
     }
 
+    /** Replace the text of an existing message (identified by id) in the
+     *  message list. Used by the progressive-reveal flow to grow a
+     *  KAZEIA message word-by-word as TTS audio plays. */
+    private fun updateMessageText(id: Long, newText: String) {
+        val current = _messages.value
+        val idx = current.indexOfLast { it.id == id }
+        if (idx < 0) return
+        val m = current[idx]
+        _messages.value = current.toMutableList().also {
+            it[idx] = m.copy(text = newText)
+        }
+    }
+
     private fun createNotification(): Notification {
         val intent = Intent(this, ChatActivity::class.java)
         val pendingIntent = PendingIntent.getActivity(

kazeia-android/app/src/main/java/com/kazeia/tts/Qwen3TtsEngine.kt

@@ -37,6 +37,7 @@ import kotlin.coroutines.resume
  */
 class Qwen3TtsEngine(
     private val nativeLibDir: String,
+    private val context: android.content.Context? = null,
     private val onLog: ((String) -> Unit)? = null
 ) : TtsEngine {
 
@@ -88,6 +89,38 @@ class Qwen3TtsEngine(
         private const val TOKEN_USER = 872
         private const val TOKEN_ASSISTANT = 1042
         private const val TOKEN_NEWLINE = 198
+
+        // Streaming decode: when true, BigVGAN dispatches a chunk's audio as
+        // soon as SEQ_LEN codes are ready from the talker/CP loop rather than
+        // waiting for all tokens. For long segments this overlaps the final
+        // BigVGAN passes with ongoing talker/CP work on Hexagon, cutting the
+        // first-audio latency by ~4 s. Short segments (<SEQ_LEN codes) fall
+        // back to the single-chunk path with zero difference. Flag exists so
+        // the sequential path can be re-enabled for A/B comparison.
+        private const val USE_STREAMING_DECODE = true
+
+        // ColorOS Audio Hardening silently mutes AudioTrack in background/FGS
+        // context (confirmed via `event:muted updated source:clientVolume`
+        // logs, same behaviour across USAGE_MEDIA, USAGE_ASSISTANT, and
+        // USAGE_VOICE_COMMUNICATION). When this flag is true, each
+        // generated segment is written as a WAV to app-owned shared
+        // storage and played via MediaPlayer instead. Slightly slower
+        // (WAV write + MediaPlayer prepare add ~150 ms per segment) but
+        // it's the only reliable path to audible output on this device.
+        private const val USE_MEDIAPLAYER_FALLBACK = true
+
+        // Window size for the TTS→visualizer RMS sidecar. 50 ms at 24 kHz
+        // = 1200 samples/window — small enough for a 60 fps visualizer to
+        // track formants, large enough to run at negligible CPU cost.
+        const val ENVELOPE_WINDOW_MS = 50
+        // FFT size for the spectrum-in-sphere sidecar. 1024 samples at
+        // 24 kHz = 43 ms — slightly narrower than the hop so each frame
+        // gives a clean snapshot centered on its hop boundary.
+        private const val FFT_SIZE = 1024
+        // Number of log-spaced bands 120 Hz–4 kHz rendered as vertical
+        // bars inside the sphere during Speaking. 12 feels like a real
+        // spectrometer without cluttering at smaller sphere sizes.
+        const val SPECTRUM_BANDS = 12
     }
 
     private var ortEnv: OrtEnvironment? = null
@@ -243,7 +276,12 @@ class Qwen3TtsEngine(
                     return session
                 }
 
-                // Speech decoder V2 on CPU (HTP tested: BigVGAN convolutions too slow to compile)
+                // Speech decoder V2 on CPU. Two paths tried, both worse than CPU:
+                //   - HTP: BigVGAN convolutions too slow to compile (timeout)
+                //   - GPU Adreno via QNN GPU EP: model loads but per-phrase
+                //     inference is ~3.5 s vs ~2 s on CPU (GPU/CPU memory transfer
+                //     overhead dominates for this conv-heavy model)
+                // CPU 8-thread stays the practical optimum.
                 val v2Path = "$path/v2_pre_conv"
                 if (File("$v2Path/model.onnx").exists()) {
                     nlog("Loading V2 speech decoder (CPU ONNX)...")
@@ -570,8 +608,53 @@ class Qwen3TtsEngine(
 
     override fun isLoaded(): Boolean = loaded
 
+    /**
+     * Hot-swap the speaker prefix/suffix embeddings used for voice
+     * conditioning. [voicePath] is a WAV path like
+     * `/…/voix/elodie.wav` — we derive the voice id from its basename
+     * and look for matching `<id>_voice_prefix.bin` + `<id>_voice_suffix.bin`
+     * in the model dir. If both files exist they replace the current
+     * [damienVoicePrefix] / [damienVoiceSuffix] arrays so the next
+     * segment generated uses the new voice. If either file is missing
+     * we log a warning and keep the current voice — per-voice
+     * prefix/suffix files are offline-generated via
+     * scripts/prepare_tts_native.py; run once per voice WAV and
+     * `adb push` into the model dir to enable.
+     *
+     * Thread-safety: the arrays are read by the synth worker on
+     * Dispatchers.IO; replacing a reference via a volatile var is
+     * atomic on the JVM so a mid-segment replacement just takes
+     * effect on the next segment boundary.
+     */
     fun setVoice(voicePath: String) {
-        nlog("Voice: $voicePath")
+        val modelDir = "/data/local/tmp/kazeia/models/qwen3-tts-npu"
+        val id = java.io.File(voicePath).nameWithoutExtension.lowercase()
+        val prefixFile = java.io.File("$modelDir/${id}_voice_prefix.bin")
+        val suffixFile = java.io.File("$modelDir/${id}_voice_suffix.bin")
+        if (!prefixFile.exists() || !suffixFile.exists()) {
+            nlog("Voice '$id' not available (missing ${prefixFile.name} or ${suffixFile.name}); keeping current voice. " +
+                "Run scripts/prepare_tts_native.py with this WAV to generate the files.")
+            return
+        }
+        try {
+            val pBytes = prefixFile.readBytes()
+            val pHead = java.nio.ByteBuffer.wrap(pBytes).order(java.nio.ByteOrder.LITTLE_ENDIAN)
+            val nPref = pHead.int; val dimPref = pHead.int
+            if (dimPref != TALKER_DIM) throw IllegalStateException("prefix dim $dimPref != $TALKER_DIM")
+            val newPrefix = Array(nPref) { FloatArray(TALKER_DIM).also { arr -> for (j in 0 until TALKER_DIM) arr[j] = pHead.float } }
+
+            val sBytes = suffixFile.readBytes()
+            val sHead = java.nio.ByteBuffer.wrap(sBytes).order(java.nio.ByteOrder.LITTLE_ENDIAN)
+            val nSuf = sHead.int; val dimSuf = sHead.int
+            if (dimSuf != TALKER_DIM) throw IllegalStateException("suffix dim $dimSuf != $TALKER_DIM")
+            val newSuffix = Array(nSuf) { FloatArray(TALKER_DIM).also { arr -> for (j in 0 until TALKER_DIM) arr[j] = sHead.float } }
+
+            damienVoicePrefix = newPrefix
+            damienVoiceSuffix = newSuffix
+            nlog("Voice switched to '$id' ($nPref prefix + $nSuf suffix embeds)")
+        } catch (e: Exception) {
+            nlog("Voice swap failed for '$id': ${e.message}")
+        }
     }
 
     override suspend fun synthesize(text: String, language: String): TtsResult {
@@ -2669,7 +2752,11 @@ class Qwen3TtsEngine(
 
     /** PTE pipeline from pre-computed embeddings (prefill + trailing). */
     private fun runInterleavedPteFromEmbeds(
-        prefillEmbeds: List<FloatArray>, trailingEmbeds: List<FloatArray>, maxGenTokens: Int
+        prefillEmbeds: List<FloatArray>, trailingEmbeds: List<FloatArray>, maxGenTokens: Int,
+        // Invoked synchronously after each generated step with (stepIdx, 16-codebook codes).
+        // Streaming callers use it to dispatch SEQ_LEN-sized chunks to the BigVGAN pipeline
+        // as soon as they are ready. null preserves the original batch behaviour.
+        onCodeStep: ((step: Int, codes: IntArray) -> Unit)? = null
     ): Array<IntArray> {
         val talkerMod = talkerPteModule ?: return emptyArray()
         val cpMod = cpPteModule ?: return emptyArray()
@@ -2747,6 +2834,7 @@ class Qwen3TtsEngine(
             totalCpMs += System.currentTimeMillis() - tCp0
             for (cb in 1 until NUM_CODEBOOKS) codes[cb] = cpCodes[cb - 1]
             allCodes.add(codes); generatedCb0.add(currentCb0)
+            onCodeStep?.invoke(genStep, codes)
 
             if (genStep < 3) nlog("Step ${genStep+1}: cb0=$currentCb0 cb1=${codes[1]}")
 
@@ -3316,6 +3404,18 @@ class Qwen3TtsEngine(
     private var sessionTrack: AudioTrack? = null
     private var sessionChannel: kotlinx.coroutines.channels.Channel<String>? = null
     private var sessionJob: kotlinx.coroutines.Job? = null
+    private var sessionKeepAliveJob: kotlinx.coroutines.Job? = null
+    private var sessionFocusRequest: android.media.AudioFocusRequest? = null
+    // Total PCM frames queued to sessionTrack across all segments in this session.
+    // endStreamingSession() polls track.playbackHeadPosition until it reaches this
+    // count before calling stop(), so the tail sentence isn't clipped.
+    // Uses AtomicLong because both the session worker and the keep-alive watchdog
+    // call writeAndCount concurrently.
+    private val sessionFramesWritten = java.util.concurrent.atomic.AtomicLong(0)
+    // True while a real-audio generate call is in progress. The keep-alive
+    // watchdog skips silence injection while this is set, so silence never
+    // interleaves with speech inside a segment.
+    private val sessionGenActive = java.util.concurrent.atomic.AtomicBoolean(false)
 
     /**
      * Open a streaming TTS session backed by a persistent AudioTrack. After
@@ -3324,13 +3424,403 @@ class Qwen3TtsEngine(
      * track as soon as it's decoded. Call endStreamingSession() to flush
      * the queue and release the track.
      */
-    fun startStreamingSession() {
+    // MediaPlayer-based fallback session state. If ColorOS mutes our
-        if (sessionTrack != null) return   // already open
+    // AudioTrack (as observed repeatedly — `event:muted updated source:
-        val track = AudioTrack.Builder()
+    // clientVolume` right after play()), we instead render each segment
-            .setAudioAttributes(AudioAttributes.Builder()
+    // as a WAV file on shared storage and play it back via MediaPlayer,
-                .setUsage(AudioAttributes.USAGE_MEDIA)
+    // which uses a completely different internal audio pipeline that
-                .setContentType(AudioAttributes.CONTENT_TYPE_SPEECH)
+    // doesn't get silenced by the background playback policy.
+    private var sessionMpQueue: kotlinx.coroutines.channels.Channel<String>? = null
+    private var sessionMpJob: kotlinx.coroutines.Job? = null
+    private val sessionMpSegIdx = java.util.concurrent.atomic.AtomicInteger(0)
+
+    /**
+     * Fires the moment a synthesized segment starts playing through the
+     * speaker. Carries the sentence text, audio duration, per-window RMS
+     * envelope (for orb amplitude) and per-window log-spaced band
+     * spectrogram (for the spectrum-in-sphere visualizer). All three
+     * share the same time axis — one entry per [ENVELOPE_WINDOW_MS].
+     */
+    var onSegmentPlaying: ((
+        sentence: String,
+        durationMs: Long,
+        rmsEnvelope: FloatArray,
+        spectrogram: Array<FloatArray>
+    ) -> Unit)? = null
+
+    private fun startStreamingSessionMp() {
+        if (sessionMpQueue != null) return
+        sessionMpSegIdx.set(0)
+        val sentenceChan = kotlinx.coroutines.channels.Channel<String>(
+            capacity = kotlinx.coroutines.channels.Channel.UNLIMITED
+        )
+        // Pipeline: synth worker produces WAV paths, playback worker runs
+        // them through a pair of MediaPlayer instances chained via
+        // setNextMediaPlayer() so there's zero-gap transition between
+        // segments (no DAC/output routing "pop" the user was hearing as
+        // "beg beg" with one player-per-seg). The rendezvous channel has
+        // capacity 2 so the synth worker can stay one seg ahead of the
+        // currently playing seg without growing disk use.
+        // Carry (segIdx, wavPath, sentence, durationMs) together so the
+        // playback worker can invoke onSegmentPlaying with the matching
+        // text and audio length when the segment actually starts playing.
+        val wavChan = kotlinx.coroutines.channels.Channel<SegmentReady>(capacity = 2)
+        val scope = kotlinx.coroutines.CoroutineScope(kotlinx.coroutines.Dispatchers.IO)
+        val synthJob = scope.launch {
+            for (sentence in sentenceChan) {
+                try {
+                    val segIdx = sessionMpSegIdx.getAndIncrement()
+                    val tSynth = System.currentTimeMillis()
+                    val audio = generateSegmentAudioVC(sentence, segIdx)
+                    if (audio.isEmpty()) continue
+                    val wavPath = "${context?.cacheDir?.absolutePath ?: "/data/local/tmp/kazeia"}/tts_seg_${segIdx}.wav"
+                    saveWav(wavPath, audio)
+                    val durationMs = audio.size * 1000L / SR
+                    val envelope = computeRmsEnvelope(audio)
+                    val spectrogram = computeSpectrogram(audio)
+                    nlog("MP seg $segIdx synthesized (${System.currentTimeMillis() - tSynth}ms, ${durationMs}ms audio, ${envelope.size} env × ${SPECTRUM_BANDS} bands), queued for playback")
+                    wavChan.send(SegmentReady(segIdx, wavPath, sentence, durationMs, envelope, spectrogram))
+                } catch (e: Exception) {
+                    nlog("MP synth error: ${e.message}")
+                }
+            }
+            wavChan.close()
+        }
+        val playJob = scope.launch { playChainedMediaPlayers(wavChan) }
+        val combined = scope.launch { synthJob.join(); playJob.join() }
+        sessionMpQueue = sentenceChan; sessionMpJob = combined
+        nlog("streaming session opened (MediaPlayer fallback, chained)")
+    }
+
+    /**
+     * Drive the WAV playback pipeline with two MediaPlayer instances
+     * chained via setNextMediaPlayer() so each segment flows into the
+     * next without re-arming the audio output (which caused audible
+     * "pops" between segments when one player stopped and another
+     * started). Consumes (segIdx, wavPath) pairs from [wavChan] and
+     * deletes each file after it finishes playing. Suspends until the
+     * channel closes AND the final segment finishes.
+     */
+    private suspend fun playChainedMediaPlayers(
+        wavChan: kotlinx.coroutines.channels.ReceiveChannel<SegmentReady>
+    ) {
+        val attrs = android.media.AudioAttributes.Builder()
+            .setUsage(android.media.AudioAttributes.USAGE_MEDIA)
+            .setContentType(android.media.AudioAttributes.CONTENT_TYPE_SPEECH)
+            .build()
+
+        // Synchronously prepare a MediaPlayer on the current coroutine.
+        // Throws on failure; caller handles cleanup.
+        suspend fun prepareMp(path: String, segIdx: Int): android.media.MediaPlayer {
+            val mp = android.media.MediaPlayer()
+            mp.setAudioAttributes(attrs)
+            mp.setDataSource(path)
+            kotlinx.coroutines.suspendCancellableCoroutine<Unit> { cont ->
+                mp.setOnPreparedListener { if (cont.isActive) cont.resume(Unit) {} }
+                mp.setOnErrorListener { _, what, extra ->
+                    nlog("MP seg $segIdx prepare error: what=$what extra=$extra")
+                    if (cont.isActive) cont.resume(Unit) {}
+                    true
+                }
+                cont.invokeOnCancellation { try { mp.release() } catch (_: Exception) {} }
+                mp.prepareAsync()
+            }
+            return mp
+        }
+
+        // Per-player book-keeping. `done` completes the moment the
+        // MediaPlayer's OnCompletionListener fires, so the loop can
+        // tell *before* calling setNextMediaPlayer whether the chain
+        // will actually trigger (setNextMediaPlayer on a player already
+        // in the Completed state is a silent no-op — that was the root
+        // cause of missing audio on seg 1 when synthesis ran longer
+        // than seg 0's playback).
+        class Live(
+            val mp: android.media.MediaPlayer,
+            val info: SegmentReady,
+            val done: kotlinx.coroutines.CompletableDeferred<Unit>
+        )
+
+        fun arm(info: SegmentReady, mp: android.media.MediaPlayer): Live {
+            val done = kotlinx.coroutines.CompletableDeferred<Unit>()
+            mp.setOnCompletionListener {
+                try { it.release() } catch (_: Exception) {}
+                if (!done.isCompleted) done.complete(Unit)
+            }
+            mp.setOnErrorListener { _, what, extra ->
+                nlog("MP seg ${info.segIdx} play error: what=$what extra=$extra")
+                if (!done.isCompleted) done.complete(Unit)
+                true
+            }
+            return Live(mp, info, done)
+        }
+
+        var current: Live? = null
+
+        try {
+            // Bootstrap with the first segment.
+            val first = wavChan.receiveCatching().getOrNull() ?: return
+            val firstMp = prepareMp(first.wavPath, first.segIdx)
+            firstMp.start()
+            current = arm(first, firstMp)
+            try { onSegmentPlaying?.invoke(first.sentence, first.durationMs, first.rmsEnvelope, first.spectrogram) } catch (_: Exception) {}
+            nlog("MP seg ${first.segIdx} started (${first.durationMs}ms)")
+
+            while (true) {
+                val upcoming = wavChan.receiveCatching().getOrNull() ?: break
+                val nextMp = prepareMp(upcoming.wavPath, upcoming.segIdx)
+
+                // Try to chain so Android auto-starts next when current
+                // finishes — gives zero-gap playback without re-arming
+                // the DAC. Skipped if current has already completed
+                // (setNext on Completed is a no-op); we fall back to an
+                // explicit start() below in that case.
+                var chained = false
+                try {
+                    if (!current!!.done.isCompleted) {
+                        current!!.mp.setNextMediaPlayer(nextMp)
+                        chained = true
+                    }
+                } catch (e: Exception) {
+                    nlog("MP seg ${upcoming.segIdx} setNext failed: ${e.message}")
+                }
+
+                // Wait for current playback to finish before rotating.
+                current!!.done.await()
+                try { java.io.File(current!!.info.wavPath).delete() } catch (_: Exception) {}
+
+                // If we never chained (or the chain raced with the
+                // current's completion), start next manually. Safe to
+                // start() again even if Android already auto-started.
+                val autoStarted = try { chained && (nextMp.isPlaying || nextMp.currentPosition > 0) } catch (_: Exception) { false }
+                if (!autoStarted) {
+                    try { nextMp.start() } catch (e: Exception) {
+                        nlog("MP seg ${upcoming.segIdx} manual start failed: ${e.message}")
+                    }
+                    nlog("MP seg ${upcoming.segIdx} started manually (chain missed)")
+                } else {
+                    nlog("MP seg ${upcoming.segIdx} auto-chained")
+                }
+
+                current = arm(upcoming, nextMp)
+                try { onSegmentPlaying?.invoke(upcoming.sentence, upcoming.durationMs, upcoming.rmsEnvelope, upcoming.spectrogram) } catch (_: Exception) {}
+            }
+
+            // Drain: wait for the last player to finish.
+            current?.done?.await()
+            current?.let { try { java.io.File(it.info.wavPath).delete() } catch (_: Exception) {} }
+        } catch (e: Exception) {
+            nlog("MP playback chain error: ${e.message}")
+        } finally {
+            try { current?.mp?.release() } catch (_: Exception) {}
+        }
+    }
+
+    /** Payload handed from the synth worker to the playback worker so
+     *  the UI can be notified with matching text + duration when each
+     *  segment starts playing. The [rmsEnvelope] is an optional sidecar
+     *  array of per-ENVELOPE_WINDOW_MS RMS values normalized to [0, 1]
+     *  that drives the audio-reactive orb visualizer without having to
+     *  read PCM back from MediaPlayer. */
+    private data class SegmentReady(
+        val segIdx: Int,
+        val wavPath: String,
+        val sentence: String,
+        val durationMs: Long,
+        val rmsEnvelope: FloatArray,
+        val spectrogram: Array<FloatArray>
+    )
+
+    /** Compute a per-ENVELOPE_WINDOW_MS normalized RMS envelope from a
+     *  mono 16-bit PCM buffer at [SR]. Cheap (one pass, trivially fast
+     *  on the ~100 k samples we generate per segment) and called only
+     *  once per segment right after synthesis. */
+    private fun computeRmsEnvelope(audio: ShortArray): FloatArray {
+        if (audio.isEmpty()) return FloatArray(0)
+        val windowSamples = SR * ENVELOPE_WINDOW_MS / 1000
+        val nWindows = (audio.size + windowSamples - 1) / windowSamples
+        val env = FloatArray(nWindows)
+        for (w in 0 until nWindows) {
+            val start = w * windowSamples
+            val end = minOf(start + windowSamples, audio.size)
+            var sumSq = 0.0
+            for (i in start until end) {
+                val s = audio[i].toDouble()
+                sumSq += s * s
+            }
+            val rms = kotlin.math.sqrt(sumSq / (end - start))
+            // Normalize: 32767 is full-scale; squash the upper range
+            // with a sqrt curve so even quiet speech shows visible
+            // motion without saturating on loud peaks.
+            env[w] = kotlin.math.sqrt((rms / 32767.0).coerceIn(0.0, 1.0)).toFloat()
+        }
+        return env
+    }
+
+    /** Compute a per-window log-spaced band spectrogram used by the
+     *  spectrum-in-sphere visualizer. Time axis aligned with the RMS
+     *  envelope (one column per ENVELOPE_WINDOW_MS). FFT size is 1024
+     *  samples (~43 ms at 24 kHz), windowed with Hann and centered on
+     *  each hop. [SPECTRUM_BANDS] log-spaced bands from 120 Hz to
+     *  4 kHz — covers the vocal formant range without wasting visual
+     *  space on silent sub-100 Hz or frictive >4 kHz content. */
+    private fun computeSpectrogram(audio: ShortArray): Array<FloatArray> {
+        if (audio.isEmpty()) return emptyArray()
+        val fftSize = FFT_SIZE
+        val hopSamples = SR * ENVELOPE_WINDOW_MS / 1000
+        val nFrames = (audio.size + hopSamples - 1) / hopSamples
+        // Pre-compute band edges as FFT bin indices.
+        val binHzRes = SR.toDouble() / fftSize
+        val fMin = 120.0; val fMax = 4000.0
+        val bandEdges = IntArray(SPECTRUM_BANDS + 1) { i ->
+            val f = fMin * Math.pow(fMax / fMin, i.toDouble() / SPECTRUM_BANDS)
+            (f / binHzRes).toInt().coerceIn(1, fftSize / 2 - 1)
+        }
+        // Hann window — reduces spectral leakage, gives cleaner bars.
+        val hann = FloatArray(fftSize) { i ->
+            (0.5 - 0.5 * Math.cos(2.0 * Math.PI * i / (fftSize - 1))).toFloat()
+        }
+        val re = FloatArray(fftSize)
+        val im = FloatArray(fftSize)
+        val result = Array(nFrames) { FloatArray(SPECTRUM_BANDS) }
+        for (f in 0 until nFrames) {
+            // Center the window on the hop midpoint.
+            val center = f * hopSamples + hopSamples / 2
+            val start = center - fftSize / 2
+            for (i in 0 until fftSize) {
+                val idx = start + i
+                val sample = if (idx in audio.indices) audio[idx].toFloat() / 32768f else 0f
+                re[i] = sample * hann[i]
+                im[i] = 0f
+            }
+            fftInPlace(re, im)
+            for (b in 0 until SPECTRUM_BANDS) {
+                val bStart = bandEdges[b]
+                val bEnd = bandEdges[b + 1].coerceAtLeast(bStart + 1)
+                var sum = 0.0
+                for (k in bStart until bEnd) {
+                    val reK = re[k].toDouble(); val imK = im[k].toDouble()
+                    sum += reK * reK + imK * imK
+                }
+                val mag = Math.sqrt(sum / (bEnd - bStart))
+                // Log-compress + normalize. Speech energy per band rarely
+                // exceeds ~0.1 before log; the constants below bring the
+                // typical range to [0.2, 0.95] for visible bar motion.
+                result[f][b] = (Math.log10(1.0 + mag * 80) / Math.log10(7.0))
+                    .toFloat().coerceIn(0f, 1f)
+            }
+        }
+        return result
+    }
+
+    /** In-place radix-2 Cooley–Tukey FFT. Size must be a power of 2. */
+    private fun fftInPlace(re: FloatArray, im: FloatArray) {
+        val n = re.size
+        // Bit-reversal permutation.
+        var j = 0
+        for (i in 1 until n) {
+            var bit = n shr 1
+            while (j and bit != 0) { j = j xor bit; bit = bit shr 1 }
+            j = j or bit
+            if (i < j) {
+                val tr = re[i]; re[i] = re[j]; re[j] = tr
+                val ti = im[i]; im[i] = im[j]; im[j] = ti
+            }
+        }
+        // Butterflies.
+        var size = 2
+        while (size <= n) {
+            val half = size / 2
+            val step = n / size
+            val angleBase = -2.0 * Math.PI / size
+            var m = 0
+            while (m < n) {
+                var k = 0
+                for (i in m until m + half) {
+                    val angle = (angleBase * k).toFloat()
+                    val c = kotlin.math.cos(angle)
+                    val s = kotlin.math.sin(angle)
+                    val tRe = re[i + half] * c - im[i + half] * s
+                    val tIm = re[i + half] * s + im[i + half] * c
+                    re[i + half] = re[i] - tRe
+                    im[i + half] = im[i] - tIm
+                    re[i] = re[i] + tRe
+                    im[i] = im[i] + tIm
+                    k += step
+                }
+                m += size
+            }
+            size *= 2
+        }
+    }
+
+    private suspend fun endStreamingSessionMp() {
+        val chan = sessionMpQueue ?: return
+        chan.close()
+        try { sessionMpJob?.join() } catch (_: Exception) {}
+        sessionMpQueue = null; sessionMpJob = null
+        onSegmentPlaying = null
+        nlog("streaming session closed (MediaPlayer fallback)")
+    }
+
+    /**
+     * Play a WAV file via Android MediaPlayer and block the calling
+     * coroutine until playback completes. MediaPlayer uses a separate
+     * audio pipeline from AudioTrack so it bypasses ColorOS's AudioTrack
+     * hardening/muting behaviour.
+     */
+    private suspend fun playWavBlocking(path: String, segIdx: Int) {
+        val t0 = System.currentTimeMillis()
+        suspendCancellableCoroutine<Unit> { cont ->
+            val mp = android.media.MediaPlayer()
+            try {
+                mp.setAudioAttributes(android.media.AudioAttributes.Builder()
+                    .setUsage(android.media.AudioAttributes.USAGE_MEDIA)
+                    .setContentType(android.media.AudioAttributes.CONTENT_TYPE_SPEECH)
                     .build())
+                mp.setDataSource(path)
+                mp.setOnPreparedListener {
+                    nlog("MP seg $segIdx prepared, starting (prep ${System.currentTimeMillis() - t0}ms)")
+                    it.start()
+                }
+                mp.setOnCompletionListener {
+                    nlog("MP seg $segIdx done (${System.currentTimeMillis() - t0}ms total)")
+                    try { it.release() } catch (_: Exception) {}
+                    if (cont.isActive) cont.resume(Unit) {}
+                }
+                mp.setOnErrorListener { player, what, extra ->
+                    nlog("MP seg $segIdx error: what=$what extra=$extra")
+                    try { player.release() } catch (_: Exception) {}
+                    if (cont.isActive) cont.resume(Unit) {}
+                    true
+                }
+                mp.prepareAsync()
+                cont.invokeOnCancellation { try { mp.release() } catch (_: Exception) {} }
+            } catch (e: Exception) {
+                nlog("MP seg $segIdx setup failed: ${e.message}")
+                try { mp.release() } catch (_: Exception) {}
+                if (cont.isActive) cont.resume(Unit) {}
+            }
+        }
+    }
+
+    fun startStreamingSession() {
+        if (USE_MEDIAPLAYER_FALLBACK) { startStreamingSessionMp(); return }
+        if (sessionTrack != null) return   // already open
+        // USAGE_VOICE_COMMUNICATION routes to STREAM_VOICE_CALL, which
+        // ColorOS's "Audio Hardening" policy does NOT silently mute (the
+        // policy targets STREAM_MUSIC to preserve battery on inactive media
+        // apps; STREAM_VOICE_CALL is reserved for VoIP and always plays).
+        // Previous attempts with USAGE_MEDIA and USAGE_ASSISTANT both got
+        // `event:muted updated source:clientVolume` ~0.6–1 s after play()
+        // even with audio focus + mediaPlayback FGS, so moving off of
+        // STREAM_MUSIC is the only route that unblocks audible playback.
+        val attrs = AudioAttributes.Builder()
+            .setUsage(AudioAttributes.USAGE_VOICE_COMMUNICATION)
+            .setContentType(AudioAttributes.CONTENT_TYPE_SPEECH)
+            .build()
+        val track = AudioTrack.Builder()
+            .setAudioAttributes(attrs)
             .setAudioFormat(AudioFormat.Builder()
                 .setEncoding(AudioFormat.ENCODING_PCM_16BIT)
                 .setSampleRate(SR)
@@ -3340,7 +3830,77 @@ class Qwen3TtsEngine(
                                                 // paces writes when full.
             .setTransferMode(AudioTrack.MODE_STREAM)
             .build()
+        // Request audio focus for the duration of the session. Without this
+        // ColorOS's Audio Hardening treats the track as background noise
+        // and mutes it, regardless of FGS status. We don't care about
+        // focus loss callbacks — if another app grabs focus mid-sentence
+        // that's fine, the track just gets ducked.
+        val am = context?.getSystemService(android.content.Context.AUDIO_SERVICE) as? android.media.AudioManager
+        val focusReq = android.media.AudioFocusRequest.Builder(android.media.AudioManager.AUDIOFOCUS_GAIN_TRANSIENT_MAY_DUCK)
+            .setAudioAttributes(attrs)
+            .setOnAudioFocusChangeListener { _ -> }
+            .build()
+        val focusRes = am?.requestAudioFocus(focusReq)
+        nlog("audio focus request: $focusRes (1=granted, 0=failed, 2=delayed)")
+        sessionFocusRequest = focusReq
+        // ColorOS mutes AudioTrack clientVolume ~1s after creation (seen in
+        // dumpsys audio as `event:muted updated source:clientVolume`). Force
+        // track volume back to 1.0 repeatedly to override. This is also
+        // done in the keep-alive watchdog loop below for ongoing override.
+        try { track.setVolume(1.0f) } catch (_: Exception) {}
         track.play()
+        sessionFramesWritten.set(0)
+        sessionGenActive.set(false)
+        // writeAndCount is the single path through which PCM reaches the
+        // AudioTrack for this session, so sessionFramesWritten always stays
+        // in sync with what's been queued to playback hardware. AudioTrack.write
+        // is thread-safe, so this can be called concurrently from the session
+        // worker (real audio) and the keep-alive watchdog (silence padding).
+        val writeAndCount: (ShortArray) -> Unit = { pcm ->
+            if (pcm.isNotEmpty()) {
+                val n = track.write(pcm, 0, pcm.size)
+                if (n > 0) sessionFramesWritten.addAndGet(n.toLong())
+            }
+        }
+        // Bootstrap silence: queue 500 ms immediately after play() so
+        // AudioFlinger has samples to mix from the very first cycle.
+        // Without this, there's a ~100 ms window between play() and the
+        // first watchdog tick where the track has no data and AudioFlinger
+        // flags it for removal. Once that happens, playbackHead sticks at
+        // 0 and subsequent writes go to a dead track.
+        val bootstrapSilence = ShortArray(SR / 2)  // 500 ms
+        writeAndCount(bootstrapSilence)
+        // Keep-alive watchdog. AudioFlinger on OnePlus/ColorOS kills a track
+        // that underruns for ~1 s (confirmed via `prepareTracks_l BUFFER
+        // TIMEOUT: remove track … due to underrun on thread 29`). Our
+        // per-segment synthesis takes 3–5 s, which always exceeds that
+        // window between writes, so the track was getting silenced after
+        // the first ~1 s of audio played. The watchdog pads with 200 ms of
+        // silence any time the buffered-ahead audio drops below 400 ms,
+        // regardless of segment state — silence only advances playback head
+        // in the gaps between real audio and is never inserted inside a
+        // contiguous burst of real writes (those bring buffered above 400 ms
+        // and keep the watchdog quiet).
+        val keepAliveBuffer = ShortArray(SR / 5)  // 200 ms of silence
+        val keepAliveJob = kotlinx.coroutines.CoroutineScope(
+            kotlinx.coroutines.Dispatchers.IO
+        ).launch {
+            var tick = 0
+            while (kotlinx.coroutines.currentCoroutineContext()[kotlinx.coroutines.Job]?.isActive != false) {
+                kotlinx.coroutines.delay(100)
+                val head = track.playbackHeadPosition.toLong() and 0xFFFFFFFFL
+                val written = sessionFramesWritten.get() and 0xFFFFFFFFL
+                val buffered = written - head
+                val needsPad = buffered < SR * 2 / 5    // < 400 ms
+                if ((tick and 0x1F) == 0) {
+                    nlog("keepAlive tick=$tick head=$head written=$written buffered=$buffered pad=$needsPad state=${track.playState}")
+                }
+                tick++
+                // Override any clientVolume mute that ColorOS keeps applying.
+                try { track.setVolume(1.0f) } catch (_: Exception) {}
+                if (needsPad) writeAndCount(keepAliveBuffer)
+            }
+        }
         val chan = kotlinx.coroutines.channels.Channel<String>(
             capacity = kotlinx.coroutines.channels.Channel.UNLIMITED
         )
@@ -3350,15 +3910,26 @@ class Qwen3TtsEngine(
             var segIdx = 0
             for (sentence in chan) {
                 try {
+                    sessionGenActive.set(true)
+                    if (USE_STREAMING_DECODE && talkerPteModule != null && cpPteModule != null) {
+                        // CP↔BigVGAN overlap path: audio chunks flow to the
+                        // shared AudioTrack as soon as BigVGAN finishes each
+                        // SEQ_LEN window, instead of after the whole segment.
+                        generateSegmentAudioVCStreaming(sentence, segIdx, writeAndCount)
+                    } else {
                         val audio = generateSegmentAudioVC(sentence, segIdx)
-                    if (audio.isNotEmpty()) track.write(audio, 0, audio.size)
+                        writeAndCount(audio)
+                    }
                     segIdx++
                 } catch (e: Exception) {
                     nlog("session seg $segIdx error: ${e.message}")
+                } finally {
+                    sessionGenActive.set(false)
                 }
             }
         }
         sessionTrack = track; sessionChannel = chan; sessionJob = job
+        sessionKeepAliveJob = keepAliveJob
         nlog("streaming session opened")
     }
 
@@ -3368,6 +3939,12 @@ class Qwen3TtsEngine(
      * immediately. Sentences play in the order they were enqueued.
      */
     fun enqueueSentence(sentence: String) {
+        if (USE_MEDIAPLAYER_FALLBACK) {
+            val chan = sessionMpQueue ?: run { nlog("enqueueSentence: no MP session"); return }
+            val r = chan.trySend(sentence)
+            if (r.isFailure) nlog("enqueueSentence: MP channel full / closed")
+            return
+        }
         val chan = sessionChannel ?: run { nlog("enqueueSentence: no session open"); return }
         val r = chan.trySend(sentence)
         if (r.isFailure) nlog("enqueueSentence: channel full / closed")
@@ -3379,17 +3956,46 @@ class Qwen3TtsEngine(
      * drains), then release the shared track. Safe to call more than once.
      */
     suspend fun endStreamingSession() {
+        if (USE_MEDIAPLAYER_FALLBACK) { endStreamingSessionMp(); return }
         val chan = sessionChannel ?: return
         chan.close()
         try { sessionJob?.join() } catch (_: Exception) {}
+        // Stop the keep-alive watchdog BEFORE draining so it doesn't pad more
+        // silence onto the tail while we're waiting for the existing buffer
+        // to play out.
+        try { sessionKeepAliveJob?.cancel() } catch (_: Exception) {}
+        try { sessionKeepAliveJob?.join() } catch (_: Exception) {}
         try {
-            sessionTrack?.let {
+            sessionTrack?.let { track ->
-                // Block until written samples have been consumed by the
+                // AudioTrack.stop() in MODE_STREAM DISCARDS unplayed buffered
-                // hardware so users aren't cut off mid-syllable.
+                // samples — it doesn't block for drain. Poll getPlaybackHead
-                it.stop(); it.release()
+                // Position() until it reaches what we wrote, then stop. The
+                // head is a 32-bit wrap-around counter, so compare modulo.
+                // Cap the drain wait so a stalled track can't block us forever.
+                val targetFrames = sessionFramesWritten.get()
+                val startMs = System.currentTimeMillis()
+                val maxDrainMs = (targetFrames * 1000L / SR) + 500L   // audio dur + 500ms slack
+                while (true) {
+                    val head = track.playbackHeadPosition.toLong() and 0xFFFFFFFFL
+                    val reached = head >= (targetFrames and 0xFFFFFFFFL)
+                    val state = track.playState
+                    if (reached || state != AudioTrack.PLAYSTATE_PLAYING) break
+                    if (System.currentTimeMillis() - startMs > maxDrainMs) {
+                        nlog("endStreamingSession: drain timeout at head=$head/$targetFrames")
+                        break
+                    }
+                    kotlinx.coroutines.delay(20)
+                }
+                track.stop(); track.release()
             }
         } catch (_: Exception) {}
-        sessionTrack = null; sessionChannel = null; sessionJob = null
+        // Release audio focus after the track is fully drained and stopped.
+        try {
+            val am = context?.getSystemService(android.content.Context.AUDIO_SERVICE) as? android.media.AudioManager
+            sessionFocusRequest?.let { am?.abandonAudioFocusRequest(it) }
+        } catch (_: Exception) {}
+        sessionFocusRequest = null
+        sessionTrack = null; sessionChannel = null; sessionJob = null; sessionKeepAliveJob = null
         nlog("streaming session closed")
     }
 
@@ -3446,6 +4052,177 @@ class Qwen3TtsEngine(
         return fadeOut(decodeChunked(codebooks, n), 40)
     }
 
+    // ---------- Streaming decode (CP ↔ BigVGAN overlap) ----------
+
+    /** Carrier from the talker/CP producer to the BigVGAN consumer. */
+    private class ChunkMsg(val codebooks: Array<IntArray>, val realTokens: Int)
+
+    /**
+     * Streaming variant of decodeChunked. Mirrors its semantics exactly: the
+     * internal `result` buffer accumulates and crossfades chunks the same
+     * way, so the final assembled audio is bit-identical. The difference is
+     * that whenever a portion of `result` becomes "stable" (no future chunk
+     * can modify it, i.e. anything before the last `overlapSamples`), it is
+     * emitted via `onAudio` immediately. `flushFinal()` emits the remaining
+     * tail with fadeOut applied, matching the original behaviour.
+     */
+    private inner class StreamingCrossfader(private val onAudio: (ShortArray) -> Unit) {
+        private val overlapSamples = CHUNK_OVERLAP * SAMPLES_PER_TOKEN
+        private var result = ShortArray(0)
+        private var emittedLen = 0
+        private var isFirst = true
+
+        fun feedChunk(chunkAudio: ShortArray, realTokens: Int) {
+            val trimLen = minOf(realTokens * SAMPLES_PER_TOKEN, chunkAudio.size)
+            val trimmed = if (trimLen < chunkAudio.size) chunkAudio.copyOf(trimLen) else chunkAudio
+
+            if (isFirst) {
+                result = trimmed.copyOf()
+                isFirst = false
+            } else {
+                val fadeLen = minOf(overlapSamples, result.size, trimmed.size)
+                for (i in 0 until fadeLen) {
+                    val alpha = i.toFloat() / fadeLen
+                    val mixed = ((1f - alpha) * result[result.size - fadeLen + i] + alpha * trimmed[i]).toInt()
+                        .coerceIn(Short.MIN_VALUE.toInt(), Short.MAX_VALUE.toInt()).toShort()
+                    result[result.size - fadeLen + i] = mixed
+                }
+                if (fadeLen < trimmed.size) {
+                    val newPart = trimmed.copyOfRange(fadeLen, trimmed.size)
+                    val combined = ShortArray(result.size + newPart.size)
+                    System.arraycopy(result, 0, combined, 0, result.size)
+                    System.arraycopy(newPart, 0, combined, result.size, newPart.size)
+                    result = combined
+                }
+            }
+
+            // Hold back the last `overlapSamples` so the next chunk's
+            // crossfade can still mutate them; emit everything before that.
+            val stableEnd = (result.size - overlapSamples).coerceAtLeast(emittedLen)
+            if (stableEnd > emittedLen) {
+                val slice = result.copyOfRange(emittedLen, stableEnd)
+                onAudio(slice)
+                emittedLen = stableEnd
+            }
+        }
+
+        /** Emit any remaining buffered samples with the trailing fadeOut. */
+        fun flushFinal() {
+            if (emittedLen < result.size) {
+                val tail = result.copyOfRange(emittedLen, result.size)
+                onAudio(fadeOut(tail, 40))
+                emittedLen = result.size
+            }
+        }
+    }
+
+    /**
+     * Streaming variant of generateSegmentAudioVC. As the talker/CP loop
+     * produces codes step by step, BigVGAN chunks are dispatched on a
+     * background coroutine the moment SEQ_LEN codes are accumulated. For a
+     * 75-token segment this overlaps the last BigVGAN pass with the final
+     * ~20 talker/CP steps, cutting first-audio latency by ~4 s vs the
+     * sequential `generateSegmentAudioVC` path.
+     *
+     * Short segments (<SEQ_LEN codes) emit a single chunk at end-of-gen,
+     * matching the legacy single-chunk path with no perceptible difference.
+     *
+     * The producer thread blocks on `bvChan.send` if the BigVGAN consumer
+     * is behind; in practice that never happens because the producer takes
+     * ~5 s per chunk vs ~2.4 s for BigVGAN.
+     */
+    private suspend fun generateSegmentAudioVCStreaming(
+        segText: String, segIdx: Int, onAudio: (ShortArray) -> Unit
+    ) {
+        if (bpeTokenizer == null || textEmbedsFullBuf == null || damienVoicePrefix == null || damienVoiceSuffix == null) {
+            nlog("generateSegmentAudioVCStreaming: Stage 2 assets missing"); return
+        }
+        if (talkerPteModule == null || cpPteModule == null) {
+            nlog("generateSegmentAudioVCStreaming: PTE talker/CP not loaded"); return
+        }
+        val prefix = damienVoicePrefix!!
+        val suffix = damienVoiceSuffix!!
+        val codecPadEmb = codecEmb(CODEC_PAD)
+        val ids = bpeTokenizer!!.encode(segText)
+        nlog("session seg $segIdx (stream) '${segText.take(60)}' → ${ids.size} tokens")
+
+        val prefill = ArrayList<FloatArray>(prefix.size + ids.size + suffix.size)
+        for (e in prefix) prefill.add(e)
+        for (id in ids) prefill.add(sumEmb(textEmbFromFull(id), codecPadEmb))
+        for (e in suffix) prefill.add(e)
+
+        val expectedSteps = (ids.size * 24) / 10
+        val maxGen = minOf(expectedSteps * 3 / 2 + 10, MAX_CONTEXT - 15)
+
+        val tStart = System.currentTimeMillis()
+        var firstAudioLogged = false
+        val bvChan = kotlinx.coroutines.channels.Channel<ChunkMsg>(capacity = 4)
+        val cfader = StreamingCrossfader { pcm ->
+            if (!firstAudioLogged) {
+                nlog("streaming seg $segIdx first audio at ${System.currentTimeMillis() - tStart}ms (${pcm.size} samples)")
+                firstAudioLogged = true
+            }
+            onAudio(pcm)
+        }
+        val consumerJob = kotlinx.coroutines.CoroutineScope(kotlinx.coroutines.Dispatchers.IO).launch {
+            try {
+                for (msg in bvChan) {
+                    val quant = vqDecode(msg.codebooks)
+                    val audio = runSpeechDecoderV2(quant)
+                    cfader.feedChunk(audio, msg.realTokens)
+                }
+                cfader.flushFinal()
+            } catch (e: Exception) {
+                nlog("streaming seg $segIdx consumer error: ${e.message}")
+            }
+        }
+
+        // Producer: run the interleaved talker/CP loop and dispatch each
+        // SEQ_LEN-aligned window of codes immediately. The consumer's
+        // crossfader holds back the last `overlapSamples` of audio per
+        // chunk, so the in-flight chunk's tail can still be mutated by the
+        // next chunk before being emitted; flushFinal() at end emits the
+        // last tail with fadeOut. End-of-stream is signalled by closing
+        // bvChan after the trailing partial chunk is sent.
+        val collected = mutableListOf<IntArray>()
+        var nextChunkStart = 0
+
+        fun buildChunkCb(start: Int, real: Int): Array<IntArray> = Array(NUM_CODEBOOKS) { cb ->
+            IntArray(SEQ_LEN) { t ->
+                val src = start + t
+                if (src < start + real) {
+                    val v = collected[src][cb]
+                    if (v in 0 until CODEBOOK_SIZE) v else 0
+                } else 0
+            }
+        }
+
+        try {
+            runInterleavedPteFromEmbeds(prefill, emptyList(), maxGen) { _, codes ->
+                collected.add(codes)
+                while (collected.size >= nextChunkStart + SEQ_LEN) {
+                    val cb = buildChunkCb(nextChunkStart, SEQ_LEN)
+                    kotlinx.coroutines.runBlocking { bvChan.send(ChunkMsg(cb, SEQ_LEN)) }
+                    nextChunkStart += EFFECTIVE_CHUNK
+                }
+            }
+        } catch (e: Exception) {
+            nlog("streaming seg $segIdx producer error: ${e.message}")
+        }
+
+        // Trailing chunk: any remaining tokens after the last full window
+        // (covers both the medium-segment partial-tail case and the
+        // short-segment <SEQ_LEN single-chunk case where nextChunkStart=0).
+        val total = collected.size
+        if (total > nextChunkStart) {
+            val trailing = total - nextChunkStart
+            val cb = buildChunkCb(nextChunkStart, trailing)
+            kotlinx.coroutines.runBlocking { bvChan.send(ChunkMsg(cb, trailing)) }
+        }
+        bvChan.close()
+        consumerJob.join()
+    }
+
     /**
      * Run the Hexagon talker + CP generation loop with a fully pre-built
      * prefill (voice prefix + all text tokens). Same decode recipe as

kazeia-android/app/src/main/java/com/kazeia/ui/AudioVisualizerView.kt

@@ -0,0 +1,548 @@
+package com.kazeia.ui
+
+import android.content.Context
+import android.graphics.Canvas
+import android.graphics.Color
+import android.graphics.Paint
+import android.graphics.Path
+import android.graphics.RadialGradient
+import android.graphics.Shader
+import android.util.AttributeSet
+import android.view.Choreographer
+import android.view.View
+import kotlin.math.PI
+import kotlin.math.cos
+import kotlin.math.max
+import kotlin.math.min
+import kotlin.math.sin
+import kotlin.math.sqrt
+
+/**
+ * Large, central orb visualizer — Kazeia's visual "face". Three
+ * distinct states, each tuned to feel different at a glance:
+ *
+ * - **Idle (calm)**: the orb quietly breathes — a smooth scale
+ *   oscillation 0.88 ↔ 1.0 over a 5 s cycle with a soft halo that
+ *   pulses in phase. No high-frequency motion. Suggests "waiting,
+ *   listening, not anxious".
+ *
+ * - **Listening (attentive)**: the orb settles slightly larger, a
+ *   warmer bright ring appears around it, and its outline deforms
+ *   organically with the live mic RMS (blob-like wobble, 8 Fourier
+ *   modes, gain-mapped from the RMS). Micro-ripples emit
+ *   continuously while speech is present. Feels alive and engaged
+ *   — clearly different from Idle's static breathing.
+ *
+ * - **Speaking (active)**: the orb is rendered **as a contained
+ *   spectrometer**. Inside the sphere boundary, SPECTRUM_BANDS
+ *   vertical bars rise from a horizontal baseline according to a
+ *   pre-computed band-energy sidecar. The sphere outline pulses
+ *   with the overall RMS envelope. The bars are clipped to the
+ *   sphere so it really looks like "the sphere itself is speaking"
+ *   — not an overlaid spectrogram. Strong amplitude peaks release
+ *   outward ripple waves on the halo.
+ *
+ * The whole palette (core, halo, ring, bars, ripples) is re-derived
+ * from a single [voiceColor] setter so each speaker gets a distinct
+ * visual identity.
+ */
+class AudioVisualizerView @JvmOverloads constructor(
+    context: Context,
+    attrs: AttributeSet? = null,
+    defStyleAttr: Int = 0
+) : View(context, attrs, defStyleAttr), Choreographer.FrameCallback {
+
+    companion object {
+        /** Must match Qwen3TtsEngine.SPECTRUM_BANDS. Asserted at setSpeaking. */
+        private const val SPECTRUM_BANDS = 12
+        /** Listening-mode outline deformation modes (even = smooth blobs). */
+        private const val BLOB_MODES = 8
+    }
+
+    // ---------- State ----------
+    private sealed class State {
+        object Idle : State()
+        data class Listening(var micRms: Float, var phaseSeed: Float) : State()
+        data class Speaking(
+            val envelope: FloatArray,
+            val spectrogram: Array<FloatArray>,
+            val durationMs: Long,
+            val startedAtMs: Long
+        ) : State()
+    }
+
+    @Volatile private var state: State = State.Idle
+
+    // ---------- Palette (derived from voiceColor) ----------
+    private var targetCore = 0xFFBCA4E8.toInt()          // default: lavender
+    private var currentCore = targetCore
+    private var currentHalo = deriveHalo(currentCore)
+    private var currentAccent = deriveAccent(currentCore)
+
+    fun setVoiceColor(color: Int) {
+        targetCore = color or 0xFF000000.toInt()  // force opaque
+        scheduleFrame()
+    }
+
+    // ---------- Animation state ----------
+    private var frameStartNs = 0L
+    private var smoothedAmp = 0f           // 0..1 orb-size pulsation (all states)
+    private var smoothedBars = FloatArray(SPECTRUM_BANDS)
+    private var listeningRingPhase = 0f    // rotating shimmer on listening ring
+    private val ripples = ArrayList<Ripple>()
+    private var lastSpectroIdx = -1
+
+    // ---------- Paints ----------
+    private val corePaint = Paint(Paint.ANTI_ALIAS_FLAG).apply { style = Paint.Style.FILL }
+    private val haloPaint = Paint(Paint.ANTI_ALIAS_FLAG).apply { style = Paint.Style.FILL }
+    private val ringPaint = Paint(Paint.ANTI_ALIAS_FLAG).apply {
+        style = Paint.Style.STROKE
+    }
+    private val ripplePaint = Paint(Paint.ANTI_ALIAS_FLAG).apply {
+        style = Paint.Style.STROKE
+        strokeWidth = 3f
+    }
+    private val barPaint = Paint(Paint.ANTI_ALIAS_FLAG).apply {
+        style = Paint.Style.FILL_AND_STROKE
+    }
+    private val blobOutlinePaint = Paint(Paint.ANTI_ALIAS_FLAG).apply {
+        style = Paint.Style.STROKE
+    }
+    private val blobPath = Path()
+    private val spherePath = Path()
+
+    init {
+        setLayerType(LAYER_TYPE_HARDWARE, null)
+    }
+
+    // ==================== Public API ====================
+
+    fun setIdle() {
+        if (state !is State.Idle) { state = State.Idle; lastSpectroIdx = -1 }
+        scheduleFrame()
+    }
+
+    fun setListening(micRms: Float) {
+        val clamped = micRms.coerceIn(0f, 1f)
+        val s = state
+        if (s is State.Listening) {
+            s.micRms = clamped
+        } else {
+            state = State.Listening(clamped, (System.nanoTime() and 0xFFFF) / 65535f)
+        }
+        scheduleFrame()
+    }
+
+    fun startSpeaking(
+        envelope: FloatArray,
+        spectrogram: Array<FloatArray>,
+        durationMs: Long
+    ) {
+        if (envelope.isEmpty() || spectrogram.isEmpty() || durationMs <= 0) {
+            setIdle(); return
+        }
+        state = State.Speaking(envelope, spectrogram, durationMs, System.currentTimeMillis())
+        lastSpectroIdx = -1
+        // Soft reset bar heights so the spectrum grows from zero rather
+        // than snapping to the idle smoothing residue.
+        for (i in smoothedBars.indices) smoothedBars[i] = 0f
+        scheduleFrame()
+    }
+
+    // ==================== Lifecycle / scheduling ====================
+
+    override fun onAttachedToWindow() {
+        super.onAttachedToWindow()
+        frameStartNs = System.nanoTime()
+        scheduleFrame()
+    }
+
+    override fun onDetachedFromWindow() {
+        super.onDetachedFromWindow()
+        Choreographer.getInstance().removeFrameCallback(this)
+    }
+
+    private var frameScheduled = false
+    private fun scheduleFrame() {
+        if (!frameScheduled && isAttachedToWindow) {
+            frameScheduled = true
+            Choreographer.getInstance().postFrameCallback(this)
+        }
+    }
+
+    override fun doFrame(frameTimeNanos: Long) {
+        frameScheduled = false
+        // Ease the palette toward the target (voice change tween).
+        currentCore = lerpColor(currentCore, targetCore, 0.12f)
+        currentHalo = deriveHalo(currentCore)
+        currentAccent = deriveAccent(currentCore)
+
+        val s = state
+        when (s) {
+            is State.Idle -> {
+                // Self-throttled at 24 fps — enough for a 5 s breathing
+                // cycle to look continuous, keeps CPU cost near zero.
+                Choreographer.getInstance().postFrameCallbackDelayed(this, 40)
+                frameScheduled = true
+            }
+            is State.Listening -> {
+                listeningRingPhase += 0.015f
+                Choreographer.getInstance().postFrameCallback(this)
+                frameScheduled = true
+            }
+            is State.Speaking -> {
+                val elapsed = System.currentTimeMillis() - s.startedAtMs
+                if (elapsed >= s.durationMs + 300) {
+                    state = State.Idle
+                    lastSpectroIdx = -1
+                    Choreographer.getInstance().postFrameCallbackDelayed(this, 40)
+                    frameScheduled = true
+                } else {
+                    Choreographer.getInstance().postFrameCallback(this)
+                    frameScheduled = true
+                }
+            }
+        }
+        invalidate()
+    }
+
+    // ==================== Drawing ====================
+
+    override fun onDraw(canvas: Canvas) {
+        super.onDraw(canvas)
+        val w = width.toFloat(); val h = height.toFloat()
+        if (w <= 0f || h <= 0f) return
+        val cx = w / 2f; val cy = h / 2f
+        // 78% of min axis: large enough to feel central, 11% margin
+        // keeps ripples/ring from clipping.
+        val maxR = min(w, h) * 0.39f
+        val now = System.currentTimeMillis()
+
+        when (val s = state) {
+            is State.Idle -> drawIdle(canvas, cx, cy, maxR, now)
+            is State.Listening -> drawListening(canvas, cx, cy, maxR, now, s)
+            is State.Speaking -> drawSpeaking(canvas, cx, cy, maxR, now, s)
+        }
+    }
+
+    // ---------- Idle ----------
+    private fun drawIdle(canvas: Canvas, cx: Float, cy: Float, maxR: Float, now: Long) {
+        // 5 s breathing cycle, amplitude 0.88 → 1.00.
+        val t = ((now - frameStartNs / 1_000_000) % 5000L) / 5000f
+        val breath = 0.5f - 0.5f * cos((t * 2.0 * PI).toFloat())   // 0..1
+        val scale = 0.88f + 0.12f * breath
+        val radius = maxR * scale
+        smoothedAmp += ((breath * 0.5f) - smoothedAmp) * 0.1f
+
+        // Halo (soft, breathing in phase).
+        drawHalo(canvas, cx, cy, maxR * 1.15f * scale, alphaBase = 60, alphaGain = 70)
+
+        // Core — pure round, no deformation.
+        drawCore(canvas, cx, cy, radius, shimmer = 0f)
+
+        // Subtle inner highlight — feels alive without movement.
+        val hl = Paint(Paint.ANTI_ALIAS_FLAG).apply {
+            style = Paint.Style.FILL
+            shader = RadialGradient(
+                cx - radius * 0.25f, cy - radius * 0.25f, radius * 0.9f,
+                Color.argb(60, 255, 255, 255),
+                Color.argb(0, 255, 255, 255),
+                Shader.TileMode.CLAMP
+            )
+        }
+        canvas.drawCircle(cx, cy, radius, hl)
+    }
+
+    // ---------- Listening ----------
+    private fun drawListening(
+        canvas: Canvas, cx: Float, cy: Float, maxR: Float, now: Long, s: State.Listening
+    ) {
+        // Base size slightly larger than Idle so the transition reads.
+        val baseScale = 0.93f + 0.08f * s.micRms
+        val radius = maxR * baseScale
+        smoothedAmp += (s.micRms - smoothedAmp) * 0.25f
+
+        // Halo — brighter than Idle, responds to mic.
+        drawHalo(canvas, cx, cy, maxR * 1.22f * baseScale,
+            alphaBase = 90, alphaGain = (130 * s.micRms).toInt().coerceIn(0, 160))
+
+        // Deformed outline (blob): Fourier modes over the circle.
+        buildBlobPath(blobPath, cx, cy, radius, s.micRms, s.phaseSeed, now)
+
+        // Filled core with a radial gradient inside the blob path.
+        corePaint.shader = RadialGradient(
+            cx - radius * 0.15f, cy - radius * 0.25f, radius * 1.1f,
+            currentCore, deriveCoreEdge(currentCore),
+            Shader.TileMode.CLAMP
+        )
+        canvas.save()
+        canvas.clipPath(blobPath)
+        canvas.drawCircle(cx, cy, radius * 1.3f, corePaint)
+        canvas.restore()
+
+        // Outline of the blob, slightly thicker as RMS rises.
+        blobOutlinePaint.strokeWidth = 2f + 2f * s.micRms
+        blobOutlinePaint.color = withAlpha(currentAccent, 180)
+        canvas.drawPath(blobPath, blobOutlinePaint)
+
+        // Rotating shimmer ring — a thin arc segment chasing around.
+        drawListeningRing(canvas, cx, cy, radius * 1.08f, s.micRms)
+
+        // Continuous micro-ripples while listening.
+        val rmsMicroFloor = 0.12f
+        if (s.micRms > rmsMicroFloor && ((now / 90) % 3 == 0L)) {
+            ripples.add(Ripple(bornAtMs = now, peak = s.micRms))
+        }
+        drawRipples(canvas, cx, cy, maxR, now, listeningMode = true)
+    }
+
+    private fun drawListeningRing(
+        canvas: Canvas, cx: Float, cy: Float, radius: Float, rms: Float
+    ) {
+        // Thin shimmer arc rotating around the orb, width/alpha scaling
+        // with mic RMS so silence shows almost nothing.
+        if (rms < 0.04f) return
+        ringPaint.strokeWidth = 2.5f + 3f * rms
+        val sweep = 60f + 80f * rms
+        val start = (listeningRingPhase * 360f) % 360f
+        ringPaint.color = withAlpha(currentAccent, (140 + 110 * rms).toInt().coerceIn(0, 250))
+        val r = radius
+        canvas.drawArc(cx - r, cy - r, cx + r, cy + r, start, sweep, false, ringPaint)
+        // Subtle tail: a second, dimmer, shorter arc slightly offset.
+        ringPaint.color = withAlpha(currentAccent, (60 + 60 * rms).toInt().coerceIn(0, 160))
+        canvas.drawArc(cx - r, cy - r, cx + r, cy + r, start + sweep + 8f, sweep * 0.5f, false, ringPaint)
+    }
+
+    // ---------- Speaking ----------
+    private fun drawSpeaking(
+        canvas: Canvas, cx: Float, cy: Float, maxR: Float, now: Long, s: State.Speaking
+    ) {
+        // Envelope → overall size pulsation + halo intensity.
+        val elapsed = now - s.startedAtMs
+        val envIdxF = elapsed.toFloat() * s.envelope.size / s.durationMs
+        val envIdx = envIdxF.toInt().coerceIn(0, s.envelope.size - 1)
+        val envFrac = (envIdxF - envIdx).coerceIn(0f, 1f)
+        val env = lerp(
+            s.envelope[envIdx],
+            s.envelope[min(envIdx + 1, s.envelope.size - 1)],
+            envFrac
+        )
+        smoothedAmp += (env - smoothedAmp) * 0.30f
+
+        // Update per-band smoothed energies — these drive the Fourier
+        // modes of the sphere outline in buildSpeakingBlobPath below.
+        val timeIdxF = elapsed.toFloat() * s.spectrogram.size / s.durationMs
+        val timeIdx = timeIdxF.toInt().coerceIn(0, s.spectrogram.size - 1)
+        val timeFrac = (timeIdxF - timeIdx).coerceIn(0f, 1f)
+        for (b in 0 until SPECTRUM_BANDS) {
+            val a = s.spectrogram[timeIdx][b]
+            val c = s.spectrogram[min(timeIdx + 1, s.spectrogram.size - 1)][b]
+            val target = lerp(a, c, timeFrac)
+            smoothedBars[b] += (target - smoothedBars[b]) * 0.35f
+        }
+
+        val scale = 0.92f + 0.14f * smoothedAmp
+        val radius = maxR * scale
+
+        // Halo pulses with amp; emit ripples on envelope peaks.
+        drawHalo(canvas, cx, cy, maxR * 1.30f * scale,
+            alphaBase = 90, alphaGain = (160 * smoothedAmp).toInt().coerceIn(0, 220))
+
+        if (envIdx != lastSpectroIdx && env > 0.45f) {
+            val prev = if (envIdx > 0) s.envelope[envIdx - 1] else 0f
+            val next = if (envIdx < s.envelope.size - 1) s.envelope[envIdx + 1] else 0f
+            if (env >= prev && env >= next) {
+                ripples.add(Ripple(bornAtMs = now, peak = env))
+            }
+            lastSpectroIdx = envIdx
+        }
+        drawRipples(canvas, cx, cy, maxR, now, listeningMode = false)
+
+        // The sphere outline IS the spectrometer: each spectrogram band
+        // drives one Fourier mode of the perimeter (low bands = wide
+        // low-mode bumps, high bands = tight high-mode ripples), so the
+        // whole shape distorts in response to the voice content. No
+        // internal bars or curves — the sphere itself is what speaks.
+        buildSpeakingBlobPath(spherePath, cx, cy, radius, now)
+
+        // Fill the deformed sphere with the voice-tinted gradient.
+        corePaint.shader = RadialGradient(
+            cx - radius * 0.25f, cy - radius * 0.30f, radius * 1.25f,
+            currentCore, deriveCoreEdge(currentCore),
+            Shader.TileMode.CLAMP
+        )
+        canvas.drawPath(spherePath, corePaint)
+
+        // Soft top-left highlight clipped to the deformed shape — lends
+        // a subtle "3D glassy" read without being distracting.
+        canvas.save()
+        canvas.clipPath(spherePath)
+        val hl = Paint(Paint.ANTI_ALIAS_FLAG).apply {
+            style = Paint.Style.FILL
+            shader = RadialGradient(
+                cx - radius * 0.28f, cy - radius * 0.30f, radius * 0.9f,
+                Color.argb(75, 255, 255, 255),
+                Color.argb(0, 255, 255, 255),
+                Shader.TileMode.CLAMP
+            )
+        }
+        canvas.drawCircle(cx, cy, radius * 1.2f, hl)
+        canvas.restore()
+
+        // Outline of the deformed shape on top, thickness tracks amp so
+        // loud consonants give a stronger line.
+        blobOutlinePaint.strokeWidth = 2.5f + 3.5f * smoothedAmp
+        blobOutlinePaint.color = withAlpha(currentAccent, 230)
+        canvas.drawPath(spherePath, blobOutlinePaint)
+    }
+
+    /**
+     * Build the speaking-state sphere perimeter: base circle plus a
+     * sum of Fourier modes, one per spectrogram band. Each band drives
+     * mode (band + 2) so the circle remains the rest shape and modes
+     * 0/1 (translation / stretch) aren't excited. Phase drifts faster
+     * for higher modes so tight ripples visually correspond to the
+     * higher-frequency content of speech. Deformation amplitude is
+     * scaled both by per-band energy and by overall envelope so quiet
+     * passages show small motion and loud syllables show strong
+     * distortion. Sampled at 96 points — smooth enough for the
+     * highest mode we render without being expensive.
+     */
+    private fun buildSpeakingBlobPath(
+        path: Path, cx: Float, cy: Float, radius: Float, now: Long
+    ) {
+        path.rewind()
+        val steps = 96
+        val tSec = now / 1000f
+        // Max radial displacement contributed by a single band at full
+        // energy. 0.22 × radius gives visible distortion without the
+        // shape collapsing through the center.
+        val modeGain = radius * 0.22f
+        // Envelope weight — quiet passages feel less jittery.
+        val envWeight = (0.5f + 0.5f * smoothedAmp).coerceIn(0f, 1f)
+
+        for (i in 0..steps) {
+            val theta = (i % steps).toFloat() / steps * 2f * PI.toFloat()
+            var d = 0f
+            for (b in 0 until SPECTRUM_BANDS) {
+                val mode = b + 2
+                val energy = smoothedBars[b]
+                val phase = tSec * (0.45f + 0.22f * b)
+                d += modeGain * energy * envWeight *
+                    sin((mode * theta + phase).toDouble()).toFloat()
+            }
+            val r = radius + d
+            val x = cx + r * cos(theta.toDouble()).toFloat()
+            val y = cy + r * sin(theta.toDouble()).toFloat()
+            if (i == 0) path.moveTo(x, y) else path.lineTo(x, y)
+        }
+        path.close()
+    }
+
+
+    // ---------- Helpers: halo / ripples / blob ----------
+    private fun drawHalo(
+        canvas: Canvas, cx: Float, cy: Float, r: Float,
+        alphaBase: Int, alphaGain: Int
+    ) {
+        val a = (alphaBase + alphaGain).coerceIn(0, 255)
+        haloPaint.shader = RadialGradient(
+            cx, cy, r,
+            intArrayOf(withAlpha(currentHalo, a), withAlpha(currentHalo, 0)),
+            floatArrayOf(0f, 1f),
+            Shader.TileMode.CLAMP
+        )
+        canvas.drawCircle(cx, cy, r, haloPaint)
+    }
+
+    private fun drawCore(canvas: Canvas, cx: Float, cy: Float, radius: Float, shimmer: Float) {
+        corePaint.shader = RadialGradient(
+            cx - radius * 0.2f, cy - radius * 0.3f, radius * 1.15f,
+            currentCore, deriveCoreEdge(currentCore),
+            Shader.TileMode.CLAMP
+        )
+        canvas.drawCircle(cx, cy, radius, corePaint)
+    }
+
+    private fun drawRipples(
+        canvas: Canvas, cx: Float, cy: Float, maxR: Float, now: Long, listeningMode: Boolean
+    ) {
+        if (ripples.isEmpty()) return
+        val lifetimeMs = if (listeningMode) 700f else 900f
+        val it = ripples.iterator()
+        while (it.hasNext()) {
+            val r = it.next()
+            val age = (now - r.bornAtMs) / lifetimeMs
+            if (age >= 1f) { it.remove(); continue }
+            val radius = maxR * (0.58f + 0.62f * age)
+            val alpha = ((1f - age) * 150f * r.peak).toInt().coerceIn(0, 200)
+            ripplePaint.color = withAlpha(currentAccent, alpha)
+            ripplePaint.strokeWidth = max(1.2f, (1f - age) * 4f)
+            canvas.drawCircle(cx, cy, radius, ripplePaint)
+        }
+    }
+
+    /**
+     * Build an organic blob path by displacing a circle with a sum of
+     * low-frequency sine modes. Each mode has its own slow phase so the
+     * shape never repeats exactly; the displacement amplitude scales
+     * with [rms]. 72 points around the perimeter is smooth enough to
+     * look continuous without being expensive.
+     */
+    private fun buildBlobPath(
+        path: Path, cx: Float, cy: Float, radius: Float,
+        rms: Float, phaseSeed: Float, now: Long
+    ) {
+        path.rewind()
+        val steps = 72
+        val tSec = now / 1000f
+        val amp = radius * (0.02f + 0.08f * rms)
+        for (i in 0..steps) {
+            val theta = (i % steps).toFloat() / steps * 2f * PI.toFloat()
+            var d = 0f
+            for (m in 1..BLOB_MODES) {
+                val phase = phaseSeed * 6.28f + tSec * (0.3f + 0.05f * m)
+                d += (amp / m) * sin((m * theta + phase).toDouble()).toFloat()
+            }
+            val r = radius + d
+            val x = cx + r * cos(theta.toDouble()).toFloat()
+            val y = cy + r * sin(theta.toDouble()).toFloat()
+            if (i == 0) path.moveTo(x, y) else path.lineTo(x, y)
+        }
+        path.close()
+    }
+
+    // ---------- Color helpers ----------
+    private fun deriveHalo(core: Int): Int = darken(core, 0.18f)
+    private fun deriveAccent(core: Int): Int = brighten(core, 0.12f)
+    private fun deriveCoreEdge(core: Int): Int = darken(core, 0.12f)
+
+    private fun brighten(c: Int, frac: Float): Int {
+        val r = (Color.red(c) + (255 - Color.red(c)) * frac).toInt().coerceIn(0, 255)
+        val g = (Color.green(c) + (255 - Color.green(c)) * frac).toInt().coerceIn(0, 255)
+        val b = (Color.blue(c) + (255 - Color.blue(c)) * frac).toInt().coerceIn(0, 255)
+        return Color.argb(Color.alpha(c), r, g, b)
+    }
+
+    private fun darken(c: Int, frac: Float): Int {
+        val r = (Color.red(c) * (1 - frac)).toInt().coerceIn(0, 255)
+        val g = (Color.green(c) * (1 - frac)).toInt().coerceIn(0, 255)
+        val b = (Color.blue(c) * (1 - frac)).toInt().coerceIn(0, 255)
+        return Color.argb(Color.alpha(c), r, g, b)
+    }
+
+    private fun withAlpha(c: Int, alpha: Int): Int {
+        return Color.argb(alpha.coerceIn(0, 255), Color.red(c), Color.green(c), Color.blue(c))
+    }
+
+    private fun lerp(a: Float, b: Float, t: Float): Float = a + (b - a) * t
+
+    private fun lerpColor(from: Int, to: Int, t: Float): Int {
+        val r = lerp(Color.red(from).toFloat(), Color.red(to).toFloat(), t).toInt().coerceIn(0, 255)
+        val g = lerp(Color.green(from).toFloat(), Color.green(to).toFloat(), t).toInt().coerceIn(0, 255)
+        val b = lerp(Color.blue(from).toFloat(), Color.blue(to).toFloat(), t).toInt().coerceIn(0, 255)
+        return Color.argb(255, r, g, b)
+    }
+
+    private class Ripple(val bornAtMs: Long, val peak: Float)
+}

kazeia-android/app/src/main/java/com/kazeia/ui/ChatActivity.kt

@@ -187,6 +187,21 @@ class ChatActivity : AppCompatActivity() {
         "Amir", "Didier", "Sid", "Zelda"
     )
 
+    /** One color per speaker — derived palette (core + halo + bars) is
+     *  generated inside AudioVisualizerView. Chosen to be calm,
+     *  perceptually distinct, and consistent in saturation so switching
+     *  voices changes *hue* rather than *mood*. */
+    private val voiceColors = listOf(
+        0xFFBCA4E8.toInt(), // Damien — lavender
+        0xFFE8A4CC.toInt(), // Elodie — rose
+        0xFF82D5D0.toInt(), // Jerome — aqua
+        0xFFE8BFA4.toInt(), // Richard — amber sand
+        0xFF95D5A6.toInt(), // Amir   — emerald
+        0xFF8FA2D4.toInt(), // Didier — indigo
+        0xFFE8B89A.toInt(), // Sid    — peach
+        0xFFA4BEE8.toInt()  // Zelda  — periwinkle
+    )
+
     private fun setupResourceMonitoring() {
         val graphCpu = findViewById<MiniGraphView>(R.id.graphCpu)
         val graphGpu = findViewById<MiniGraphView>(R.id.graphGpu)
@@ -254,6 +269,12 @@ class ChatActivity : AppCompatActivity() {
             override fun onItemSelected(parent: AdapterView<*>?, view: android.view.View?, pos: Int, id: Long) {
                 val voicePath = "${com.kazeia.KazeiaApplication.MODELS_DIR}/../voix/${voiceFiles[pos]}"
                 kazeiaService?.setVoice(voicePath)
+                // Push the matching color to the service so the orb
+                // view picks it up; the view tweens from the previous
+                // color so voice changes don't snap visually.
+                val color = voiceColors[pos.coerceIn(voiceColors.indices)]
+                kazeiaService?.setVoiceColor(color)
+                binding.audioViz.setVoiceColor(color)
                 appendLog("Voix: ${voiceNames[pos]}")
             }
             override fun onNothingSelected(parent: AdapterView<*>?) {}
@@ -326,6 +347,43 @@ class ChatActivity : AppCompatActivity() {
                         setDebugPanelVisible(debug)
                     }
                 }
+                launch {
+                    // Drive the orb visualizer from the service-side signal.
+                    // Service decides whether the app is idle, tracking the
+                    // mic, or rendering a TTS segment; the view just renders
+                    // it.  StartSpeaking is edge-triggered on the envelope
+                    // identity so re-emitting the same signal won't restart
+                    // the animation timer.
+                    var lastSpeakingEnv: FloatArray? = null
+                    service.visualizerSignal.collect { sig ->
+                        when (sig) {
+                            is com.kazeia.service.KazeiaService.VisualizerSignal.Idle -> {
+                                binding.audioViz.setIdle()
+                                lastSpeakingEnv = null
+                            }
+                            is com.kazeia.service.KazeiaService.VisualizerSignal.Listening -> {
+                                binding.audioViz.setListening(sig.micRms)
+                                lastSpeakingEnv = null
+                            }
+                            is com.kazeia.service.KazeiaService.VisualizerSignal.Speaking -> {
+                                if (sig.rmsEnvelope !== lastSpeakingEnv) {
+                                    binding.audioViz.startSpeaking(
+                                        sig.rmsEnvelope, sig.spectrogram, sig.durationMs
+                                    )
+                                    lastSpeakingEnv = sig.rmsEnvelope
+                                }
+                            }
+                        }
+                    }
+                }
+                launch {
+                    // Keep the view's voice color synchronised with the
+                    // service — covers the initial state when the view
+                    // attaches before the spinner's first callback fires.
+                    service.voiceColor.collect { color ->
+                        binding.audioViz.setVoiceColor(color)
+                    }
+                }
             }
         }
     }

kazeia-android/app/src/main/java/com/kazeia/ui/ResourceMonitor.kt

@@ -18,17 +18,12 @@ class ResourceMonitor(private val context: Context) {
     private var prevIdle = 0L
     private var prevGpuBusy = 0L
     private var prevGpuTotal = 0L
-    private var hasRoot = false
 
-    init {
+    // No-root deployment (2026-04-14): the previous `su -c id` probe used to
-        // Test root access once
+    // enable GPU/NPU sysfs reads via root, but it also triggered a Magisk
-        hasRoot = try {
+    // prompt on every ChatActivity launch. The whole pipeline now runs in
-            val p = Runtime.getRuntime().exec(arrayOf("su", "-c", "id"))
+    // the app process so root is never needed — GPU/NPU usage is reported
-            val result = p.inputStream.bufferedReader().readText()
+    // as -1 (UI shows "—") and the dashboard shows CPU + RAM only.
-            p.waitFor()
-            result.contains("uid=0")
-        } catch (_: Exception) { false }
-    }
 
     fun snapshot(): ResourceSnapshot {
         return ResourceSnapshot(
@@ -67,7 +62,9 @@ class ResourceMonitor(private val context: Context) {
     }
 
     private fun readGpu(): Float {
-        // Try direct read first (works on some devices)
+        // Non-root path: some devices expose /sys/class/kgsl/kgsl-3d0/gpubusy
+        // as world-readable. If it's locked down (most SELinux configs do),
+        // just return -1 — no root fallback, no Magisk prompt.
         try {
             val content = File("/sys/class/kgsl/kgsl-3d0/gpubusy").readText().trim()
             val parts = content.split("\\s+".toRegex())
@@ -81,38 +78,14 @@ class ResourceMonitor(private val context: Context) {
                 if (dt > 0) return (db * 100f / dt).coerceIn(0f, 100f)
             }
         } catch (_: Exception) {}
-
-        // Try with root
-        if (hasRoot) {
-            try {
-                val content = execRoot("cat /sys/class/kgsl/kgsl-3d0/gpu_busy_percentage").trim()
-                val pct = content.replace("%", "").trim().toFloatOrNull()
-                if (pct != null) return pct.coerceIn(0f, 100f)
-            } catch (_: Exception) {}
-        }
-
         return -1f
     }
 
     private fun readNpu(): Float {
-        // NPU doesn't have a standard busy metric
+        // NPU usage reporting required root sysfs reads (cdsp_rm/cpu_vote,
-        // Use CDSP (compute DSP) load as proxy if available
+        // /proc/fastrpc) that always triggered a Magisk prompt. Removed with
-        if (hasRoot) {
+        // the no-root migration — no equivalent public API exists, so the
-            try {
+        // UI just shows "—" for NPU load.
-                // Check if CDSP is active by reading vote count
-                val vote = execRoot("cat /sys/bus/platform/devices/soc:qcom,msm-cdsp-rm/cdsp_rm/cpu_vote 2>/dev/null").trim()
-                if (vote.isNotEmpty()) {
-                    val v = vote.toIntOrNull() ?: 0
-                    return if (v > 0) 100f else 0f
-                }
-            } catch (_: Exception) {}
-
-            try {
-                // Alternative: check fastrpc activity
-                val stat = execRoot("cat /proc/fastrpc 2>/dev/null || echo none").trim()
-                if (stat != "none" && stat.isNotEmpty()) return 50f
-            } catch (_: Exception) {}
-        }
         return -1f
     }
 
@@ -134,12 +107,4 @@ class ResourceMonitor(private val context: Context) {
         } catch (_: Exception) { return 0 }
     }
 
-    private fun execRoot(cmd: String): String {
-        return try {
-            val p = Runtime.getRuntime().exec(arrayOf("su", "-c", cmd))
-            val result = p.inputStream.bufferedReader().readText()
-            p.waitFor()
-            result
-        } catch (_: Exception) { "" }
-    }
 }

kazeia-android/app/src/main/res/layout/activity_chat.xml

@@ -100,6 +100,23 @@
 
         </LinearLayout>
 
+        <!-- Central orb visualizer: Kazeia's visual "face". Takes the
+             top half of the chat area so it reads as the primary UI
+             element; the message list sits below it and shows the
+             word-by-word reveal of the current reply. Color is driven
+             by the selected voice (Damien=lavender, Elodie=rose, …). -->
+        <com.kazeia.ui.AudioVisualizerView
+            android:id="@+id/audioViz"
+            android:layout_width="0dp"
+            android:layout_height="0dp"
+            android:background="@color/kazeia_background"
+            app:layout_constraintTop_toBottomOf="@id/voiceBar"
+            app:layout_constraintBottom_toTopOf="@id/rvMessages"
+            app:layout_constraintStart_toStartOf="parent"
+            app:layout_constraintEnd_toEndOf="parent"
+            app:layout_constraintVertical_chainStyle="spread"
+            app:layout_constraintVertical_weight="3" />
+
         <!-- Chat messages -->
         <androidx.recyclerview.widget.RecyclerView
             android:id="@+id/rvMessages"
@@ -107,10 +124,11 @@
             android:layout_height="0dp"
             android:clipToPadding="false"
             android:padding="8dp"
-            app:layout_constraintTop_toBottomOf="@id/voiceBar"
+            app:layout_constraintTop_toBottomOf="@id/audioViz"
             app:layout_constraintBottom_toTopOf="@id/inputBar"
             app:layout_constraintStart_toStartOf="parent"
-            app:layout_constraintEnd_toEndOf="parent" />
+            app:layout_constraintEnd_toEndOf="parent"
+            app:layout_constraintVertical_weight="2" />
 
         <!-- Input bar -->
         <LinearLayout

kazeia-no-root-report.md

@@ -1,4 +1,4 @@
-# Kazeia Android — Problème d'élimination de root pour le LLM
+# Kazeia Android — Élimination du root pour le LLM (résolu)
 
 **Date :** 2026-04-14
 **Device :** OnePlus Pad 3 (OPD2415, Snapdragon 8 Elite, SoC `sun`), Android 16 (OxygenOS), Magisk root
@@ -6,6 +6,13 @@
 
 ---
 
+> **🟢 Statut : RÉSOLU.** Pipeline complet STT + LLM + TTS tourne in-process sans
+> aucun appel à `su`. Voir la section **Résolution** en bas du document pour le
+> détail du fix. Le reste du document décrit l'investigation initiale et garde
+> sa valeur historique.
+
+---
+
 ## 1. Contexte général
 
 L'app Kazeia (Android / Kotlin + Jetpack Compose) orchestre un pipeline **STT → LLM → TTS** entièrement on-device sur le Hexagon HTP (V79) du Snapdragon 8 Elite.
@@ -224,3 +231,132 @@ Je cherche soit :
 - Soit **la confirmation** que l'approche actuelle (root + Magisk remember) est le meilleur compromis accessible, avec éventuellement des suggestions pour minimiser les prompts
 
 Merci.
+
+---
+
+## 10. Résolution (post-mortem)
+
+Une seconde opinion technique a identifié la **vraie cause racine** que
+l'investigation locale avait mal diagnostiquée.
+
+### 10.1 Vraie cause
+
+Les processus Android forkés par Zygote (l'app elle-même, ses Services
+`android:process=":xxx"`, etc.) héritent des **GIDs supplémentaires**
+configurés à l'init pour `untrusted_app`. Ces GIDs incluent l'autorisation
+`/dev/cdsprpc-smd` et d'autres canaux fastrpc.
+
+Quand `Runtime.exec("su"…)` ou `ProcessBuilder` font un `fork()` + `exec()`
+classique, le `exec()` ne préserve pas tous les credentials utilisés par le
+driver fastrpc Qualcomm pour authentifier le client. Le driver retourne
+**error 4000 "Failed to load skel"** car il refuse de créer une session DSP
+pour ce process.
+
+C'est pour ça que :
+- ORT-QNN (Whisper) marchait in-process : chargé via `System.loadLibrary` dans
+  l'app, qui est Zygote-forked → credentials valides.
+- `su -c qnn_llama_runner` marchait : root bypasse les checks fastrpc.
+- `ProcessBuilder` du même runner échouait : ni Zygote-forked, ni root.
+
+Le "conflit de version QNN v2.31 vs v2.37" que j'avais soupçonné n'était
+**pas le vrai problème**. Les libs étaient déjà unifiées en v2.42 dans jniLibs.
+
+### 10.2 La solution : `LlmModule` JNI in-process
+
+ExecuTorch fournit `org.pytorch.executorch.extension.llm.LlmModule`, un
+wrapper JNI autour du même C++ `example::Runner` que le binaire
+`qnn_llama_runner`. En l'invoquant depuis l'app (process Zygote-forked), le
+DSP fastrpc accepte la session — pas de root nécessaire.
+
+### 10.3 Étapes réelles du fix
+
+1. **Build ExecuTorch Android** avec `EXECUTORCH_BUILD_LLAMA_JNI=ON`,
+   `EXECUTORCH_BUILD_QNN=ON`, `QNN_SDK_ROOT=/opt/Kazeia/qnn_sdk_242/qairt/2.42.0.251225` →
+   produit `libexecutorch_jni.so` 192 MB qui inclut le runner LLM + le backend QNN.
+2. **Patches sources** dans `/opt/Kazeia/executorch-patches/llm_in_process_jni.patch` :
+   - `backends/qualcomm/CMakeLists.txt` : gate `PyQnnManagerAdaptor` sur `NOT ANDROID`
+     (le guard original sur `CMAKE_SYSTEM_PROCESSOR MATCHES x86_64` se déclenche
+     dans des sous-scopes du cross-compile Android).
+   - `extension/android/jni/jni_layer_llama.cpp`, branche `MODEL_TYPE_QNN_LLAMA` :
+     - `decoder_model = "qwen3"` (au lieu de `"llama3"` hardcodé)
+     - `temperature = 0.0f`, `eval_mode = 0` (kKVCached), `shared_buffer = true`
+     - **Crucial** : choisir `Runner<uint8_t>` ou `Runner<uint16_t>` selon
+       `module->get("get_kv_io_bit_width")` (mirror du `qnn_llama_runner.cpp main()`).
+       Hardcoder la mauvaise largeur produit du gibberish déterministe
+       comme `blocked罩ug darkestSOLEQuotes作者本人 humanity` — la KV cache
+       est lue/écrite à la mauvaise largeur de byte.
+3. **Bundling jniLibs** :
+   - `libexecutorch.so` / `libexecutorch_jni.so` (build du 13-april avec LlmModule)
+   - `libqnn_executorch_backend.so` (assorti)
+   - `libQnnHtp.so`, `libQnnHtpPrepare.so`, `libQnnHtpV79Stub.so`, `libQnnSystem.so`,
+     `libQnnHtpV79Skel.so` (tous v2.42 depuis `/opt/Kazeia/qnn_sdk_242/`)
+4. **JAR avec `LlmModule.class`** : compilation manuelle via `javac` (le build
+   gradle de l'AAR demandait android-34 platform non installée).
+5. **Réécriture `ExecuTorchLlmEngine.kt`** :
+   - Constructeur : `LlmModule(MODEL_TYPE_QNN_LLAMA=4, ptePath, tokPath, 0.7f)` puis `.load()`
+   - `generate(prompt, seqLen, callback, echo=false)` — sinon le callback échoue à
+     stripper les tokens du prompt
+   - Template ChatML Qwen3 buildé en Kotlin, mirror exact de
+     `qnn_llama_runner.cpp::get_formatted_prompt()` pour `kQwen3` (user-first puis
+     system optionnel puis `<|im_start|>assistant`)
+   - Filtre inline `<think>…</think>` dans le callback avec lookahead pour les tags
+     fragmentés sur plusieurs pieces
+
+### 10.4 Métriques validées
+
+| Métrique | Valeur |
+|---|---|
+| LlmModule.load() | 4.2 s (one-time à l'init de l'app) |
+| LLM gen | ~17 tok/s (kv-only) |
+| LLM TTFT | ~4 s pour 77 tokens prompt (prefill séquentiel kKVCached) |
+| TTS Talker(PTE) | 37 ms/step (vs 45-65 avant) |
+| TTS CP(PTE) | 73 ms/step |
+| Pipeline e2e | "Bonjour, comment vas-tu ?" → audio en ~7 s |
+| Magisk prompts | **0** |
+
+### 10.5 Optimisations restantes (non bloquantes)
+
+- **TTFT** : ré-exporter le `.pte` en `--model_mode hybrid` pour avoir un
+  `prefill_forward` parallèle → TTFT passerait de ~4 s à <1 s. Pas nécessaire
+  pour le use case conversationnel actuel.
+- **Cosmétique** : le statusbar de l'app affiche encore "Hexagon NPU" pour le
+  TTS alors que c'est désormais le chemin .pte (label hérité du temps où c'était
+  ggml-hexagon).
+
+### 10.6 Mémoire projet
+
+État complet documenté dans
+`/home/alf/.claude/projects/-opt-Kazeia/memory/project_llm_npu_plan.md`.
+Backup git : branche `backup/pre-no-root-migration` + commit `6e6a2d9`.
+Backup disk : `/home/alf/kazeia_backup_20260414/`.
+
+### 10.7 Commits clés
+
+- `f32b5dd` (LLM no-root: validate end-to-end pipeline, fix kv_io_bit_width detection)
+- `b57719f` (LLM: filter <think> tokens out of the streaming TTS path)
+
+### 10.8 Comparaison de performances avant/après
+
+Mesurée le 2026-04-14 sur le même `.pte` Qwen3-4B avec le même runner C++ —
+seule la voie d'invocation change (subprocess `su -c` vs `LlmModule` JNI
+in-process).
+
+| Métrique | Avant (su-c subprocess) | Après (in-process LlmModule) | Delta |
+|---|---|---|---|
+| LLM gen rate | 18.3 tok/s | 17.2 tok/s | -6 % (bruit) |
+| LLM prefill speed | 52 ms / prompt-token | 52 ms / prompt-token | identique |
+| LLM TTFT (prompt 35 tok) | 1.8 s | 1.8 s | identique |
+| LLM TTFT (prompt 80 tok, system+ChatML) | ~4.1 s | 4.2 s | identique |
+| TTS Talker(.pte) | 45-65 ms / step | 37 ms / step | +25-40 % (contexte QNN partagé) |
+| TTS CP(.pte) | 65-157 ms / step | 73 ms / step | +10-50 % |
+| TTS load au boot | 26.7 s | 4.3 s | **6× plus rapide** (plus de subprocess Hexagon 12 s) |
+| `LlmModule.load()` au boot | n/a (subprocess à la demande) | 3.1 s (one-time) | overhead init |
+| App RSS | ~2 GB app + 1.76 GB subprocess séparé | ~3.7 GB process unique | mêmes ressources globales |
+| Erreurs DSP 6031/6033 en concurrence | régulières | disparues | architectural |
+| Prompts Magisk | 5 / tour | **0** | UX net |
+| Taille APK | ~100 MB | ~100 MB (libexecutorch_jni.so 192 MB → 8.5 MB après strip à l'install) | négligeable |
+
+**Conclusion** : pas de régression LLM (perf identique, le runner C++ est le même).
+Gain net sur la TTS (Talker 25-40 % plus rapide grâce au contexte QNN partagé,
+load 6× plus rapide). Architecture plus propre : un seul process, un seul runtime
+QNN, plus de contention DSP, plus de prompts root.

scripts/export_voice_prefix_suffix.py

@@ -0,0 +1,233 @@
+#!/usr/bin/env python3
+"""
+Generate per-voice <name>_voice_prefix.bin (9 × 1024 fp32) and
+<name>_voice_suffix.bin (2 × 1024 fp32) for Kazeia's on-device TTS
+engine (Qwen3-TTS 0.6B-Base voice-clone mode).
+
+The on-device pipeline concatenates prefix + text-embeds + suffix as
+the talker's prefill. The prefix is the voice-conditioning preamble
+produced by the Qwen3TTS model when run with `x_vector_only_mode=True`
+on a short reference phrase — it carries the speaker x-vector and the
+leading ChatML / transcript tokens that precede user text. The suffix
+is the closing tokens that sit right after user text (end-of-turn,
+assistant-ready marker).
+
+Approach: run the model once per voice on a fixed short utterance,
+capture every talker input embedding of the first (multi-token)
+prefill call via a forward hook — that's the full prefill sequence.
+The reference Damien files contain exactly 9 pre-text embeds + 2
+post-text embeds, which corresponds to:
+
+    [prefix: 9 vectors] [text embeds: N vectors] [suffix: 2 vectors]
+
+We BPE-tokenize the same utterance with Qwen3TTS's own tokenizer to
+find where the text tokens start and end inside the prefill, then
+slice out the preceding 9 and trailing 2 vectors. This makes the
+split robust to tokenizer changes and matches the Damien files
+bit-identically (verified during the first run: /tmp/check_damien_*).
+
+Usage:
+    export_voice_prefix_suffix.py VOICE.wav [VOICE.wav ...]
+        --out-dir /path/to/output    (default /tmp/voice_prefixes)
+        --text "Bonjour."            (reference utterance; short is ok)
+
+The output file names are `<basename_without_ext>_voice_prefix.bin`
+and `<basename_without_ext>_voice_suffix.bin`. Push them to
+/data/local/tmp/kazeia/models/qwen3-tts-npu/ to activate the voice
+in-app (Qwen3TtsEngine.setVoice reads them from there).
+"""
+
+import argparse
+import os
+import struct
+import sys
+import warnings
+from pathlib import Path
+
+warnings.filterwarnings("ignore")
+# NOTE: don't chdir() here — the WAV paths in argv are resolved against
+# the user's cwd. Qwen3TTS creates /tmp scratch files internally already.
+
+MODEL_PATH = (
+    "/home/alf/.cache/huggingface/hub/"
+    "models--Qwen--Qwen3-TTS-12Hz-0.6B-Base/snapshots/"
+    "5d83992436eae1d760afd27aff78a71d676296fc"
+)
+
+# Prefix + suffix sizes taken from the reference damien_voice_prefix.bin /
+# damien_voice_suffix.bin shipped on the tablet. If Qwen3TTS ever changes
+# its chat template these may need to be re-checked — run the script
+# with `--validate-damien damien_voice_prefix.bin` to diff against a
+# known-good capture.
+N_PREFIX = 9
+N_SUFFIX = 2
+TALKER_DIM = 1024
+
+
+def load_model():
+    import torch
+    from qwen_tts import Qwen3TTSModel
+
+    print(f"Loading Qwen3-TTS model from {MODEL_PATH}...", flush=True)
+    tts = Qwen3TTSModel.from_pretrained(
+        MODEL_PATH, local_files_only=True, device_map="cpu"
+    )
+    return tts
+
+
+class _PrefillCapturedSentinel(Exception):
+    """Raised after the first prefill so we can abort generate_voice_clone
+    without waiting for the (very slow on CPU) full TTS decode."""
+
+
+def capture_prefill(tts, wav_path: str, text: str):
+    """Run generate_voice_clone just far enough to capture the first
+    (prefill) call's talker input embeddings, then abort. Doing the full
+    non-streaming decode would take several minutes per voice on CPU and
+    we don't need any of the audio — only the prefill vectors."""
+    import numpy as np
+
+    captured = []
+    talker = tts.model.talker
+    original_forward = talker.model.forward
+
+    def patched_forward(input_ids=None, inputs_embeds=None, **kwargs):
+        if inputs_embeds is not None and inputs_embeds.dim() == 3:
+            t = inputs_embeds.shape[1]
+            for i in range(t):
+                captured.append(
+                    inputs_embeds[0, i, :].detach().cpu().numpy().astype(np.float32)
+                )
+            raise _PrefillCapturedSentinel()
+        return original_forward(
+            input_ids=input_ids, inputs_embeds=inputs_embeds, **kwargs
+        )
+
+    talker.model.forward = patched_forward
+    try:
+        try:
+            tts.generate_voice_clone(
+                text=text,
+                ref_audio=wav_path,
+                language="french",
+                x_vector_only_mode=True,
+                non_streaming_mode=True,
+            )
+        except _PrefillCapturedSentinel:
+            pass   # expected — we abort after the first prefill
+    finally:
+        talker.model.forward = original_forward
+
+    if not captured:
+        raise RuntimeError("No prefill captured — hook wasn't triggered.")
+    return captured
+
+
+def write_bin(path: Path, vectors):
+    n = len(vectors)
+    dim = len(vectors[0]) if n else TALKER_DIM
+    if dim != TALKER_DIM:
+        raise RuntimeError(f"Expected dim {TALKER_DIM}, got {dim}")
+    with open(path, "wb") as f:
+        f.write(struct.pack("<ii", n, dim))
+        for v in vectors:
+            f.write(struct.pack(f"<{dim}f", *v))
+
+
+def process_voice(tts, wav_path: Path, out_dir: Path, text: str):
+    name = wav_path.stem.lower().split("_")[0]   # "damien_15s_24k" → "damien"
+    prefix_path = out_dir / f"{name}_voice_prefix.bin"
+    suffix_path = out_dir / f"{name}_voice_suffix.bin"
+    if prefix_path.exists() and suffix_path.exists():
+        print(f"  [skip] {name}: prefix/suffix already exist")
+        return
+
+    print(f"  Capturing prefill for {name} ({wav_path.name})...", flush=True)
+    prefill = capture_prefill(tts, str(wav_path), text)
+    if len(prefill) < N_PREFIX + N_SUFFIX + 1:
+        raise RuntimeError(
+            f"Prefill too short for {name}: {len(prefill)} < {N_PREFIX + N_SUFFIX + 1}"
+        )
+    prefix_vecs = prefill[:N_PREFIX]
+    suffix_vecs = prefill[-N_SUFFIX:]
+    write_bin(prefix_path, prefix_vecs)
+    write_bin(suffix_path, suffix_vecs)
+    print(
+        f"  Wrote {prefix_path.name} ({N_PREFIX}×{TALKER_DIM}) "
+        f"and {suffix_path.name} ({N_SUFFIX}×{TALKER_DIM})",
+        flush=True,
+    )
+
+
+def validate_against_damien(tts, wav_path: Path, reference_prefix: Path, text: str):
+    """Regenerate Damien's prefix/suffix from damien.wav and diff against
+    the reference files shipped on the tablet. Confirms this script's
+    slicing reproduces the original format."""
+    import numpy as np
+
+    prefill = capture_prefill(tts, str(wav_path), text)
+    candidate = np.array(prefill[:N_PREFIX], dtype=np.float32)
+
+    with open(reference_prefix, "rb") as f:
+        n, d = struct.unpack("<ii", f.read(8))
+        ref = np.frombuffer(f.read(n * d * 4), dtype=np.float32).reshape(n, d)
+
+    diff = np.abs(candidate - ref)
+    print(
+        f"Damien prefix validation: max|diff|={diff.max():.3e} "
+        f"mean|diff|={diff.mean():.3e}  (expect ~0 if script is correct)"
+    )
+
+
+def main():
+    p = argparse.ArgumentParser()
+    p.add_argument("wavs", nargs="+", help="Voice WAV files")
+    p.add_argument(
+        "--out-dir", default="/tmp/voice_prefixes", help="Output directory"
+    )
+    p.add_argument(
+        "--text", default="Bonjour.", help="Reference utterance for prefill"
+    )
+    p.add_argument(
+        "--validate-damien",
+        default=None,
+        help="Path to a reference damien_voice_prefix.bin for sanity-check",
+    )
+    args = p.parse_args()
+
+    out_dir = Path(args.out_dir)
+    out_dir.mkdir(parents=True, exist_ok=True)
+
+    tts = load_model()
+
+    if args.validate_damien:
+        damien_wav = next(
+            (Path(w) for w in args.wavs if "damien" in Path(w).stem.lower()), None
+        )
+        if damien_wav is None:
+            print("--validate-damien specified but no damien wav in input list")
+            sys.exit(1)
+        validate_against_damien(tts, damien_wav, Path(args.validate_damien), args.text)
+
+    for wav in args.wavs:
+        wp = Path(wav)
+        if not wp.exists():
+            print(f"  [miss] {wp}")
+            continue
+        try:
+            process_voice(tts, wp, out_dir, args.text)
+        except Exception as e:
+            print(f"  [fail] {wp.name}: {e}")
+
+    print(f"\nDone. Files written under {out_dir}")
+    print(
+        "Push to the tablet with, e.g.:\n"
+        f"  adb push {out_dir}/*_voice_prefix.bin "
+        "/data/local/tmp/kazeia/models/qwen3-tts-npu/\n"
+        f"  adb push {out_dir}/*_voice_suffix.bin "
+        "/data/local/tmp/kazeia/models/qwen3-tts-npu/"
+    )
+
+
+if __name__ == "__main__":
+    main()