Model | GitHub
Flow-matching based Japanese TTS model (approximately 3.29B parameters). Generates speech from text, optional reference audio, and optional style caption.
- Reference audio: Optional. One or more clips can be concatenated in the displayed order, up to the model's 120-second limit.
- Caption: Optional style prompt for emotion, tone, speaking style, or acoustic scene.
- Duration: By default, v4-Large predicts the output duration automatically. Use Duration Scale for Large adjustments or Seconds for exact manual control.
For longer references, multiple clean, shorter clips from the same speaker are recommended. This matches v4-Large training. A single uninterrupted long recording is accepted but has not been evaluated.