search

mct10 / RepCodec

Models and code for RepCodec: A Speech Representation Codec for Speech Tokenization
Other
147 stars 10 forks source link

Are you planning to release a training code ? #4

Closed rishikksh20 closed 7 months ago

rishikksh20 commented 9 months ago

Hi @mct10 , RepCodec performance on Whisper features looks good and I am highly impressive by your approach, so I am planning to train RepCodec on non-english languages using Whisper's feature. Do you have any timeline when to release a training code for this repo?

Thanks

mct10 commented 9 months ago

Hi, thanks for your interests! We are now waiting for the decision of our paper, which will be released on Jan 15. FYI, our training code actually is very similar to AudioDec. You can have a look if interested. Thanks.

rishikksh20 commented 9 months ago

Hi @mct10 image

generator_loss = vq_loss + 45 * reconstruction_loss getting these on tensorboard, perplexity keeps on increasing. Is this behavior being normal?

mct10 commented 9 months ago

Hi, I think it is normal. I got similar behaviors.

rishikksh20 commented 9 months ago

ok thanks

yaya-sy commented 8 months ago

Hi, any news on the release of the official training script?

mct10 commented 8 months ago

Hi, we are working on it and plan to release in about 2 weeks.

rishikksh20 commented 7 months ago

finally done