KABURI-TTS
KABURI-TTS generates both speakers of a Japanese conversation together. It learns timing from LLM-jp-Zoom1 and can reproduce overlaps, pauses and short listener responses.
Research release
The current timing pipeline differs from the paper
The current version uses separate realizer and gap models. The paper's single timing predictor remains available through paper-mode and the paper-release-v1 code tag. Speaker reference audio is not included.KABURI-TTS official model card
Preserve the two speaker channels
Practical guidance. Use the channel separation when editing dialogue or studying overlapping speech. Mixing both channels too early would remove information that the model was designed to preserve.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Model identifiers | kaburi-tts |
| Acoustic ancestry | Irodori-TTS-500M-v2, adapted for dialogue. The 500M label applies to that component.KABURI-TTS official model card |
| Text component | A fine-tuned LLM-jp-3 440M model rewrites dialogue into spoken style.KABURI-TTS official model card |
| Output | 48 kHz stereo, decoded with Semantic-DACVAE-Japanese-32dim.KABURI-TTS official model card |
| License | Apache 2.0 for the release; MIT terms also apply to the acoustic ancestor and codec weights.KABURI-TTS official model card |
Related models
- LLM-jp Moshi v1
LLM-jp / NII LLMC · LLM-jp Moshi
A Japanese full-duplex speech-dialogue derivative of Kyutai's Moshi model.
Launch date not confirmedResearch releaseDownloadable - LLM-jp-3 440M Base
LLM-jp / NII LLMC · LLM-jp-3
LLM-jp-3 440M Base is a pretrained text model with 4,096-token context and Apache 2.0 weights.
Launch date not confirmedResearch releaseDownloadable