Matcha Zh Tw En 8k GGUF

--- license: apache-2.0 language: - zh - en pipeline_tag: text-to-speech tags: - matcha-tts - tts - ggml - gguf - rapidspeech - jetson - edge base_model: Luigi/matcha-zh-tw-en-8k library_name: rapidspeech.cpp --- # matcha-zh-tw-en-8k-gguf — Matcha-TTS (CFM) zh-TW/en @ 8 kHz, **ggml/GGUF** for RapidSpeech.cpp A **GGUF** conversion of [**Luigi/matcha-zh-tw-en-8k**](https://huggingface.co/Luigi/matcha-zh-tw-en-8k) (a Taiwan-Mandarin + English, 8 kHz Matcha-TTS — CFM acoustic model + Vocos vocoder) for the **ggml** runtime [**RapidSpeech.cpp**](https://github.com/vieenrose/RapidSpeech.cpp). It lets Matcha-TTS run on **ggml-CUDA / ggml-CPU** — in particular on the **Jetson Nano gen1** (Maxwell sm_53, CUDA 10.2), where the ggml engine fits the 4 GB budget that ONNX Runtime's CUDA EP overflows. ## Contents - `matcha8k.gguf` (~105 MB) — **474 learned tensors** (acoustic 394 + Vocos 80) + all hyper-parameters (`sample_rate=8000`, `ode_steps=3`, …) packed into GGUF KV. Converted with `scripts/convert_matcha_onnx_to_gguf.py` from the source bundle's `model-steps-3.onnx` + `vocos-8khz-univ.onnx`. **Round-trip validated**: every GGUF tensor's bytes == the ONNX numpy (PASS). ## Validated on real Jetson Nano gen1 (sm_53 Maxwell, CUDA 10.2, 4 GB) Cross-compiled (x86 → aarch64, gcc-8.3 + CUDA-10.2 / kreier patch set) and run on the device: | Backend | warm RTF | audio | notes | |---|---|---|---| | RapidSpeech **ggml-CPU** | **0.507** | 3.18 s synth, valid (rms 0.118) | | | RapidSpeech **ggml-CUDA** | **0.184** | identical audio | `Tegra X1 cc 5.3`, `MATCHA_USE_CUDA=1`; ~2.75× over CPU | Real-utterance check (sentence `這個星期的研究進度。`, 10 phoneme ids): **2.40 s @ 8 kHz, 151 mel frames** — matches the ONNX reference. ggml-CUDA peaks well under the 4 GB budget (the engine's whole reason on this device). ## Usage (RapidSpeech.cpp) The GGUF self-identifies as `arch=matcha-tts`. The CFM acoustic + Vocos vocoder + iSTFT run end-to-end; the **text→phoneme frontend is supplied externally** (e.g. sherpa-onnx's Matcha frontend, `--debug=1` dumps the int32 phoneme ids). Feed the ids via the `MATCHA_IDS` env (little-endian int32 file): ```sh # CPU matcha_e2e_test matcha8k.gguf out.wav # built-in fixture, or: MATCHA_IDS=ids.bin matcha_e2e_test matcha8k.gguf out.wav # ggml-CUDA (Maxwell sm_53+; matcha defaults to CPU because CUDA-graph replay needs cc>=800) MATCHA_USE_CUDA=1 MATC...