Skip to content

pockettts.md

Latest commit

 

History

History
50 lines (35 loc) · 1.33 KB

pockettts.md

File metadata and controls

50 lines (35 loc) · 1.33 KB
Raw

PocketTTS Engine

PocketTTSEngine wraps Kyutai Labs' Pocket TTS package. The source describes it as a lightweight CPU-oriented English model with optional voice cloning.

Install

Install RealtimeTTS with the PocketTTS extra:

pip install "realtimetts[pockettts]"

Minimal Use

from RealtimeTTS import TextToAudioStream, PocketTTSEngine


if __name__ == "__main__":
    engine = PocketTTSEngine(voice="alba")
    stream = TextToAudioStream(engine)
    stream.feed("Hello from Pocket TTS.")
    stream.play()

Voice cloning:

from RealtimeTTS import PocketTTSVoice, PocketTTSEngine

voice = PocketTTSVoice(name="custom", audio_prompt_path="reference.wav")
engine = PocketTTSEngine(voice=voice)

Source Notes

  • Built-in source voices are alba, marius, javert, jean, fantine, cosette, eponine, and azelma.
  • PocketTTSVoice(name, audio_prompt_path=None) treats a built-in name without a prompt as a built-in voice.
  • The model is loaded with TTSModel.load_model().
  • Voice states are cached by voice name and prompt path.
  • Output is mono 16-bit PCM at the model sample rate, falling back to 24000 Hz.

Troubleshooting

  • pocket-tts is not installed: install pocket-tts in the active environment.
  • Unknown voice errors mean the name is not built in and no prompt WAV was supplied.