Whisper chunking algorithm increases WER

[asusdisciple]

System Info

So I made a few experiments with whisper and seamless m4tv2 on the fleurs (concatenated files to 5min samples) dataset. I used the batching functionality by setting chunk_length_s to 30s and as is turns out the WER increases by 20% over all languages compared to long form transcription (sequentially going through each file). Do you have the same behaviour? Is this a bug or expected to happen because of the chunking? 20% seems to be far too much from my point of view.

Who can help?

No response

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

Just use the default pipeline implementation of whisper on files which are a few minutes long. Its far worse when chunking is enabled.

Expected behavior

I would expect the same transcription quality or maybe a few % less but 20% is far from that.

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.