sarulab-speech

/

DialogueSidon

like 11

Follow

SaruLab Speech group 64

Hi all,

The project is very cool, especially the SSL-VAE latent space and diffusion-based predictor design.

I saw the released weights include ssl_encoder.pt2, diffusion_head.pt2, vae_decoder.pt2, and metadata.json, but it looks like the stage 1 SSL-VAE encoder used to compute clean target latents z1 and z2 is not included.

Would you be open to sharing the stage 1 SSL-VAE encoder weights, or any recommended way to reproduce the same VAE-32 latent space? We are interested in fine-tuning the latent predictor, so having the exact encoder would be very helpful.

Thanks again for releasing the project.

Best,
Richard