Skip to content

sopro.md

Latest commit

 

History

History
61 lines (47 loc) · 2.21 KB

sopro.md

File metadata and controls

61 lines (47 loc) · 2.21 KB
Raw

SoproTTS Engine

SoproTTSEngine wraps SoproTTS. It is a local reference-audio engine with a native streaming API, making it one of the more practical low-latency cloning paths in the Zaphod notes.

Install

pip install "realtimetts[sopro]"

For CUDA, install a matching PyTorch and torchaudio build before or after the Sopro package, then run pip check.

Minimal Use

from RealtimeTTS import TextToAudioStream, SoproTTSEngine, SoproTTSVoice


if __name__ == "__main__":
    voice = SoproTTSVoice(ref_audio_path="reference.wav")
    engine = SoproTTSEngine(device="cuda", voice=voice)
    stream = TextToAudioStream(engine)
    stream.feed("Hello from SoproTTS.")
    stream.play()
    engine.shutdown()

Source Notes

  • SoproTTSVoice accepts ref_audio_path, ref_audio, or prompt_wav_path. The source ignores ref_text and language.
  • Defaults include model_name="samuel-vitorino/sopro", device="cuda", max_frames=400, top_p=0.9, temperature=1.05, anti_loop=True, ref_seconds=12.0, and chunk_frames=6.
  • Passing cache_dir sets both HF_HOME and HF_HUB_CACHE before model load, because the Sopro Mimi codec uses Hugging Face environment cache settings.
  • Reference audio is prepared and cached by path, mtime, and ref_seconds.
  • Output is mono 16-bit PCM at Sopro's TARGET_SR.

Zaphod Dev-Log Notes

  • Working install used sopro==1.5.0, torch==2.11.0+cu128, torchaudio==2.11.0+cu128, transformers==5.8.0, huggingface_hub==1.14.0, and soundfile==0.13.1.
  • A sandboxed first load failed when Hugging Face was unreachable; retry with approved network passed.
  • NLTK punkt_tab needed to be copied into the Sopro venv for local smoke tests.
  • Default chunk_frames=6 gave good TTFA/abort balance in the benchmark notes; chunk_frames=3 improved TTFA but worsened RTF and abort jitter.
  • Official notes and Zaphod listening notes both warn that cloning quality can be inconsistent; listen to samples before making it a default.

Troubleshooting

  • If offline runs still call Hugging Face, set both HF_HOME and HF_HUB_CACHE or pass cache_dir.
  • If startup reports missing NLTK punkt_tab, install or copy that tokenizer data into the active venv.