howitry

Some questions about flow

4

This is an amazing open source project! Why is the flow decoder used in cosyvoice much larger than the matcha-TTS decoder? What is the purpose of increasing the decoder? Is...

CER Performance of Reconstructed Audio

6

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 the reconstructed...

Mel or wav？

1

Great work! I want to ask if you have tried using mel as input? If mel is used as input and the same bitrate is maintained (e.g. frameshift=256, encoder downsampled...

**Describe the bug** What the bug is, and how to reproduce, better with screenshots(描述bug以及复现过程，最好有截图) 当使用IterablePackingDataset时，在template.encode()阶段如果创建稍大一点的tensor（比如：debug_tensor=torch.zeros(2,100000)），就会一直卡在tensor创建阶段。 如果使用PackingDataset就没有问题，使用流式的EncodePreprocessor()也没有问题。 **Your hardware and system info** Write your system info like CUDA version/system/GPU/torch version here(在这里给出硬件信息和系统信息，如CUDA版本，系统，GPU型号和torch版本等)...

Are there any plans to optimize the fetcher_state in StatefulDataLoader?

4

Since `_IterableDatasetFetcher` has no state attribute: https://github.com/pytorch/pytorch/blob/v2.6.0/torch/utils/data/_utils/fetch.py#L19, and the current `fetcher_state:dataset_iter_state` is None: https://github.com/meta-pytorch/data/blob/v0.11.0/torchdata/stateful_dataloader/worker.py#L277, could this cause prefetched data to be discarded during resume?