Skip to content

faster-qwen.md

Latest commit

 

History

History
59 lines (45 loc) · 1.82 KB

faster-qwen.md

File metadata and controls

59 lines (45 loc) · 1.82 KB
Raw

Faster Qwen Engine

FasterQwenEngine wraps faster-qwen3-tts for local Qwen voice cloning. It is a CUDA-oriented engine that benefits from cached speaker embeddings.

Install

pip install "realtimetts[qwen]"

The source notes that the instruct parameter needs faster-qwen3-tts 0.2.5. If that version is not available on PyPI in your environment, install the upstream package from Git.

Minimal Use

from RealtimeTTS import TextToAudioStream, FasterQwenEngine, FasterQwenVoice


if __name__ == "__main__":
    voice = FasterQwenVoice(
        name="demo",
        ref_audio="reference.wav",
        ref_text="Exact transcript of the reference audio.",
        language="English",
        speaker_pt="speaker.pt",
    )
    engine = FasterQwenEngine(device="cuda", voice=voice)
    stream = TextToAudioStream(engine)
    stream.feed("Hello from Faster Qwen.")
    stream.play()

Source Notes

  • Default model is Qwen/Qwen3-TTS-12Hz-0.6B-Base.
  • A FasterQwenVoice requires either speaker_pt or both ref_audio and ref_text.
  • The engine extracts and caches the speaker embedding, then primes the underlying model cache with a sentinel key.
  • Defaults include chunk_size=2, xvec_only=True, and attn_implementation="sdpa".
  • Output is mono 16-bit PCM at 24000 Hz.

Zaphod Dev-Log Notes

  • Faster Qwen was run in isolated engine environments because the shared venv became unstable across heavy model stacks.
  • One benchmark path used a cached Hugging Face snapshot and offline mode.
  • Later Zaphod tuning recorded the local cached Qwen snapshot path and exposed Qwen tuning fields in CLI/help output.

Troubleshooting

  • If first synthesis is slow, speaker extraction and CUDA graph warmup may be happening.
  • If instruct does not work, check the installed faster-qwen3-tts version.