Thanks for nice repository.
In VALL-E, it is important that there is same speakers voice in a batch when you train NAR model.
However, you used DynamicBucketingSampler. When we use it, the data is sorted by duration. Therefore, the batch is constructed from different speakers voices.
Sorry, you are excellent.
By reading carefully your code, I understand what you did in valle.py.
In a single audio, you split it to ref and target in forward function.(when training) Thank you.
Thanks for nice repository. In VALL-E, it is important that there is same speakers voice in a batch when you train NAR model. However, you used DynamicBucketingSampler. When we use it, the data is sorted by duration. Therefore, the batch is constructed from different speakers voices.
valle/data/datamodule.py