Code Release Timeline

[hertz-pj]

This is an amazing job. Is there an estimated time for the full open source release of the code and checkpoints?

[julian-parker]

We're aiming for before the end of the month, if no unforeseen obstacles appear.

[hertz-pj]

Has there been an attempt to use Stable-Codec in the training of downstream LLM-Based TTS models? I am very curious about the impact of a vocabulary size exceeding 40k on LLM-Based TTS models.

[julian-parker]

We're hoping to release a bit more info about that along with the weights, as we've done some work on maximising this. The post-hoc residual FSQ gives quite a few possibilities for reducing vocab size though. You can run 2 codebooks with 16k vocab, or even 4 with ~700 vocab.

[hertz-pj]

Thank you for your reply, I'm looking forward to it.