metadata
license: cc-by-nc-4.0
DialogueSidon
Two-speaker dialogue separation model based on diffusion with a VAE-32 latent space.
Components
| File | Description |
|---|---|
ssl_encoder.pt2 |
w2v-BERT 2.0 backbone + latent projection heads |
diffusion_head.pt2 |
DiffusionTransformerHead — single denoising step |
vae_decoder.pt2 |
DAC VAE decoder: latents → 24 kHz audio |
metadata.json |
Latent normalisation stats, scheduler config, model dims |
Contributors
- Wataru Nakata
- Yuki Saito
- Kazuki Yamauchi
- Emiru Tsunoo
- Hiroshi Saruwatari