howitry
howitry
推理阶段的hubert+vq不是用text+参考音频的离散ssl自回归预测出来的吗？推理阶段生成出来的hubert+vq不应该就想包含参考音色吗，为什么会少音色泄露？
> we use speech tokenizer, which means we must use flow model to reconstruct the mel sequence I understand that flow is used to transform code to mel. But the...
> > When using the 40 tokens/s configuration, although the quality of the reconstructed audio is very good, there are often some mispronunciations. Have you measured the CER performance of...
> Hey @byi8220, I'll review this PR, sorry for the wait ! Are you still up to tweak/finish this PR with my guidance ? If so, if you can first...
> we keep track of number of batches yielded, so the batches that were prefetched by the workers but not yielded by the dataloader, they are fetched again and yielded...
> [@ramanishsingh](https://github.com/ramanishsingh) can clarify on how we don't "lose data" and we resume on correct data. > > But regarding saving actual data in ckpt, that is not feasible because...
> split_dataset_by_node Hello, I have some questions about your intended use: (1) It seems unnecessary to use interleaving for a single dataset. (2) For multiple datasets, it seems possible to...