Skip to content

omnivoice.md

Latest commit

 

History

History
58 lines (43 loc) · 1.6 KB

omnivoice.md

File metadata and controls

58 lines (43 loc) · 1.6 KB
Raw

OmniVoice Engine

OmniVoiceEngine wraps k2-fsa/OmniVoice for local multilingual voice cloning. It requires a reference audio file and exact transcript.

Install

pip install "realtimetts[omnivoice]"

Install a compatible PyTorch build for your CUDA or CPU environment. The source import error suggests:

pip install torch torchaudio omnivoice

Minimal Use

from RealtimeTTS import TextToAudioStream, OmniVoiceEngine, OmniVoiceVoice


if __name__ == "__main__":
    voice = OmniVoiceVoice(
        name="demo",
        ref_audio="reference.wav",
        ref_text="Exact transcript of the reference audio.",
        language="en",
    )
    engine = OmniVoiceEngine(voice=voice)
    stream = TextToAudioStream(engine)
    stream.feed("Hello from OmniVoice.")
    stream.play()

Source Notes

  • Default model is k2-fsa/OmniVoice.
  • Default device_map is cuda:0 and default dtype is torch.float16.
  • num_steps_schedule defaults to [12, 32]; the first sentence uses the first value and later sentences reuse later values.
  • preprocess_prompt and postprocess_output are forwarded to generation.
  • Output is mono 16-bit PCM at 24000 Hz.

Zaphod Dev-Log Notes

  • Isolated venv smoke tests found OmniVoice needed the emotional_wavs reference assets.
  • Later speed tuning did not promote OmniVoice above the faster local paths in that environment.

Troubleshooting

  • OmniVoiceEngine requires a voice configuration: pass an OmniVoiceVoice with both ref_audio and ref_text.
  • If CUDA memory is tight, reduce steps or test CPU only for correctness first.