Kazeia
/
kazeia
3
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Commit Graph

  • db281002d9 scripts: export per-voice prefix/suffix embeddings main Kazeia Team 2026-04-15 00:09:23 +0200
  • c2f7859dfe UI+TTS: voice hot-swap + typing dots + emoji stripping Kazeia Team 2026-04-14 23:55:07 +0200
  • b5b13780f7 UI: whole-sphere Fourier-mode deformation during speech Kazeia Team 2026-04-14 23:47:30 +0200
  • 2fe46e0f15 Fix seg-2 audio dropout + switch spectrum from bars to Bézier lines Kazeia Team 2026-04-14 23:42:43 +0200
  • 06dcd76dcb UI: large central orb w/ spectrum-inside + per-voice palette Kazeia Team 2026-04-14 23:33:38 +0200
  • 8939c680b2 UI: épuré audio-reactive orb visualizer — replaces 3D avatar for MVP Kazeia Team 2026-04-14 23:20:15 +0200
  • f17131aefb UI: reveal Kazeia reply in sync with TTS audio (per-sentence, per-word) Kazeia Team 2026-04-14 22:58:18 +0200
  • 6a958c1a10 Revert MemoryOptimizer — reclaim wasn't worth the footprint Kazeia Team 2026-04-14 22:46:58 +0200
  • 751e3e0868 memory: periodic sweep + expand kill list (photos, calendar, contacts, vending, tachyon…) Kazeia Team 2026-04-14 22:44:09 +0200
  • 39babcb158 TTS+audio+memory: ColorOS playback fixes + kill-background reclaim Kazeia Team 2026-04-14 22:37:20 +0200
  • 0632db1ee0 UI: drop Magisk prompt — ResourceMonitor stops probing su Kazeia Team 2026-04-14 18:35:18 +0200
  • 10fd10fd90 TTS: overlap CP↔BigVGAN — first audio 14.5s → 10.9s per segment Kazeia Team 2026-04-14 16:22:15 +0200
  • 67de8d4767 LLM: enable hybrid-mode export via num_sharding=1 — TTFT 2.9s → 113ms Kazeia Team 2026-04-14 15:08:31 +0200
  • a41619ed67 TTS: keep BigVGAN on CPU after GPU regression; LLM filter strips more tags Kazeia Team 2026-04-14 13:48:37 +0200
  • f4b15a72a7 LLM JNI: auto-detect eval_mode from .pte methods (kv-only vs hybrid) Kazeia Team 2026-04-14 12:45:10 +0200
  • 3d435f9cdd LLM: trim system prompt to drop ~27 prefill tokens (-1.3s TTFT) Kazeia Team 2026-04-14 12:16:11 +0200
  • 7dc6704e95 docs: add before/after performance comparison to no-root report Kazeia Team 2026-04-14 11:37:15 +0200
  • 6c7746c5d0 docs: add post-mortem to no-root report — issue resolved Kazeia Team 2026-04-14 11:19:27 +0200
  • b57719fa5e LLM: filter <think> tokens out of the streaming TTS path Kazeia Team 2026-04-14 11:16:08 +0200
  • f32b5ddfdd LLM no-root: validate end-to-end pipeline, fix kv_io_bit_width detection Kazeia Team 2026-04-14 11:11:23 +0200
  • 809a6d4fed LLM no-root: migrate to in-process LlmModule (JNI) — zero su calls Kazeia Team 2026-04-14 10:39:50 +0200
  • 6e6a2d9f82 Baseline before no-root migration: working state with root LLM backup/pre-no-root-migration Kazeia Team 2026-04-14 08:19:36 +0200
  • 364016b7b8 LLM+TTS: short-response system prompt, PTE streaming fallback Kazeia Team 2026-04-14 00:17:08 +0200
  • 9930bfa392 LLM: enable Qwen3-4B NPU (21 tok/s) in service pipeline Kazeia Team 2026-04-13 23:00:25 +0200
  • 19f934af25 LLM NPU: Qwen3-4B QNN export patches + deployment notes Kazeia Team 2026-04-13 22:56:42 +0200
  • f548e02283 TTS: dynamic EOS-rank boost terminates generation cleanly across voices Kazeia Team 2026-04-13 14:13:04 +0200
  • c25040a780 TTS: conditional tail-trim + export script accepts voice path arg Kazeia Team 2026-04-13 11:32:33 +0200
  • 0833d1bd21 TTS: route all synthesizeAndPlay calls through Stage 3 streaming session Kazeia Team 2026-04-13 11:12:14 +0200
  • 2f07901ff3 TTS Stage 3: LLM stream → sentence split → TTS session → shared AudioTrack Kazeia Team 2026-04-13 10:52:46 +0200
  • 7f1a44c23d TTS Stage 2: on-device voice-cloning TTS for arbitrary text Kazeia Team 2026-04-13 10:12:09 +0200
  • 5e416713ce TTS Stage 1 streaming: play each segment the moment it's decoded Kazeia Team 2026-04-13 08:43:30 +0200
  • de878ddf5c TTS tremor investigation: identify cross-arch numerical floor, gate diag flags Kazeia Team 2026-04-13 00:15:14 +0200
  • ee186e9049 Auto-segmentation for long texts + dynamic pipeline Kazeia Team 2026-04-10 00:08:59 +0200
  • 199bc4fbc9 Full native C++ TTS validated on short + long phrases Kazeia Team 2026-04-09 23:51:05 +0200
  • dafbe2a52b FULL NATIVE C++ TTS pipeline — any text, perfect quality Kazeia Team 2026-04-09 23:39:06 +0200
  • 09d36f2025 Root cause found + on-device embed capture + KV=100 restored Kazeia Team 2026-04-09 23:00:37 +0200
  • 3dcf73aa38 Restore KV=100 + fix as-is embeds + multi-segment support Kazeia Team 2026-04-09 22:26:20 +0200
  • 10a3904d7d Multi-segment TTS for long text: split → generate → concatenate Kazeia Team 2026-04-09 14:34:05 +0200
  • 24157c0a68 Fix: use pre-computed embeds as-is (no double codec_sum) Kazeia Team 2026-04-09 14:10:23 +0200
  • f6df1738c5 Add prepare_tts_embeds.py for any text + codec_sum fix Kazeia Team 2026-04-09 14:05:42 +0200
  • 173606dae7 Stable: decoder 8T optimization + restore pre-computed embeds Kazeia Team 2026-04-09 13:42:02 +0200
  • 42bbb96fd8 Optimize decoder: BigVGAN 8T, small models 4T → RTF 1.26 Kazeia Team 2026-04-09 13:00:05 +0200
  • a688edc9ec Reduce talker KV_LEN 100→64: saves 148ms (RTF 1.31) Kazeia Team 2026-04-09 12:47:30 +0200
  • 4dcc4bb8b3 Fix KV buffer + revert HTP decoder (BigVGAN too complex for HTP) Kazeia Team 2026-04-09 12:37:50 +0200
  • 985fd9cff9 Direct output→input KV copy: RTF 1.51 → 1.31 Kazeia Team 2026-04-09 12:23:45 +0200
  • 14f7e5b05f Optimize CP+talker: eliminate prepare_input_tensors per step Kazeia Team 2026-04-09 12:16:38 +0200
  • e647911329 Shared Module C++ pipeline: RTF 1.6 with perfect quality Kazeia Team 2026-04-09 12:05:58 +0200
  • 38c0e9874a Disable C++ pipeline (QNN non-deterministic), keep Java RTF 1.8 Kazeia Team 2026-04-09 11:42:49 +0200
  • 439629c9bf Revert "Pre-allocate Tensor/EValue in Java pipeline: 16s → 8.9s (RTF 1.9)" Kazeia Team 2026-04-09 11:03:52 +0200
  • 0f027c5fde Pre-allocate Tensor/EValue in Java pipeline: 16s → 8.9s (RTF 1.9) Kazeia Team 2026-04-09 10:59:13 +0200
  • 8e536094df Fix C++ pipeline eos/pad + disable for quality (keep Java default) Kazeia Team 2026-04-09 10:53:19 +0200
  • 3b01302cfb Fix missing eos/pad embeddings in native C++ pipeline Kazeia Team 2026-04-09 10:35:05 +0200
  • 393ce79eb5 Native C++ pipeline: RTF 1.4 (was 3.6 in Java) Kazeia Team 2026-04-09 10:09:32 +0200
  • fb6045a635 Pre-load CP heads + GPU decoder test (reverted) + headArgmaxOffset Kazeia Team 2026-04-09 09:57:01 +0200
  • 6e6c562d53 Add DSP warmup + fix pipeline thread contention Kazeia Team 2026-04-09 09:24:18 +0200
  • 8bfe6c7445 Add NEON SIMD heads argmax for CP — 2.3× speedup Kazeia Team 2026-04-09 08:55:20 +0200
  • 389ffa7c61 Initial commit: Kazeia TTS pipeline on NPU via ExecuTorch Kazeia Team 2026-04-09 08:42:11 +0200