Skip to content

moss-tts.md

Latest commit

 

History

History
89 lines (71 loc) · 3.23 KB

moss-tts.md

File metadata and controls

89 lines (71 loc) · 3.23 KB
Raw

MOSS-TTS Engine

MossTTSEngine wraps MOSS-TTS-Nano ONNX and torch runtimes. The source keeps the Realtime variant behind a clear error until a bounded local smoke exists.

Install

Install RealtimeTTS with the MOSS extra. The extra includes the shared PyPI-resolvable runtime dependencies RealtimeTTS can declare, but MOSS-TTS-Nano itself, local model/runtime assets, and CUDA choices still need deliberate setup.

pip install "realtimetts[moss]"

If you keep a local MOSS-TTS-Nano checkout, installing it editable remains a good development workflow:

pip install -e third_party/MOSS-TTS-Nano

Choose one runtime path:

  • ONNX: install the MOSS ONNX runtime dependencies and onnxruntime or onnxruntime-gpu.
  • Torch: install the MOSS torch runtime dependencies and a compatible PyTorch and torchaudio build.

Minimal Use

from RealtimeTTS import TextToAudioStream, MossTTSEngine, MossTTSVoice


if __name__ == "__main__":
    voice = MossTTSVoice(
        name="demo",
        prompt_audio_path="reference.wav",
        prompt_text="Exact transcript of the reference audio.",
    )
    engine = MossTTSEngine(voice=voice, backend="onnx", execution_provider="cpu")
    stream = TextToAudioStream(engine)
    stream.feed("Hello from MOSS TTS.")
    stream.play()
    engine.shutdown()

Source Notes

  • Supported source variant is variant="nano" with backend="onnx" or backend="torch".
  • variant="realtime" or backend="realtime" currently raises an import error.
  • Defaults target OpenMOSS-Team/MOSS-TTS-Nano and OpenMOSS-Team/MOSS-Audio-Tokenizer-Nano.
  • Default audio is 48000 Hz stereo; get_stream_info() returns channel count from the loaded runtime metadata.
  • ONNX CUDA on Windows checks MOSS_CUDA_DLL_DIRS, ORT_CUDA_DLL_DIRS, CUDA_PATH, CUDA_HOME, and CUDNN_PATH, then tries common NVIDIA install directories and onnxruntime.preload_dlls.
  • The wrapper writes temporary runtime output under test_outputs/moss_tts_internal by default.

Zaphod Dev-Log Notes

  • MOSS-TTS-Nano package metadata observed Python >=3.10, torch==2.7.0, torchaudio==2.7.0, transformers==4.57.1, and onnxruntime>=1.20.0.
  • Working Nano venv used torch==2.7.0+cu128, torchaudio==2.7.0+cu128, and onnxruntime-gpu.
  • soundfile was needed because the first Nano smoke had no working torchaudio.load() backend for the reference WAV.
  • NLTK punkt and punkt_tab were installed locally for sentence splitting.
  • Torch Nano rejects prompt_text in voice_clone; the wrapper records transcript data for reproducibility but does not pass it to that torch path.
  • ONNX CPU, ONNX CUDA, and Torch CUDA smokes all produced valid files in the dev-log, but ONNX CUDA had instability and Torch CUDA had slower-than-realtime RTF despite good TTFA.
  • The MOSS-TTS-Realtime probe used torch==2.9.1+cu128, torchaudio==2.9.1+cu128, and torchcodec==0.8.1, but full smoke attempts timed out or hit torchcodec/FFmpeg DLL issues.

Troubleshooting

  • For ONNX CUDA on Windows, set CUDNN_PATH or the explicit DLL directory env vars if CUDA/cuDNN DLLs are not found.
  • If Realtime variant is requested, switch back to variant="nano" until the Realtime path is implemented and locally validated.