# TTS sur Kazeia-Engine
Talker (talker_f32.gguf 1.78GB) + CP = transformers Qwen GGUF -> ggml-cpu engine, Talker prefill HTP possible. Decoder = libtts_decoder_ggml (chraac) inchange. Tokenizer Qwen3BpeTokenizer. STT Whisper reste ORT-QAIRT (RTF0.4 NPU, gguf=CPU only regression). Unif LLM+TTS = 1 runtime ggml, RAM partagee.