BiCodec training apparently never converges.

[Respaired]

Greetings.
I tried training the BiCodec on more than 1K hours of multilingual data on the default 32khz config. after 10K-20K steps, it results in an intelligible output but it never improves in terms of acoustic quality or speaker similarity beyond that (I have waited for two days and 72K steps).

here's the log:

20250801_102602.log

here's a sample at step 72,000:
rec_0.wav , ground_truth.wav
I appreciate your input.
Thanks!

[manhcuong17072002]

@Respaired
Hello,

May I ask you a quick question?

• After that, did you continue training, and did the results improve?
• Did you train the model from scratch, or did you fine-tune from an existing checkpoint?

I’m also interested in training a bicodec on a new language—would you mind sharing some of your experiments?

Thank you.

Edit:
Could you share your training configuration file? Also, could you provide the training duration, and any other additional information?