Pad audio instead of mel features to reduce word error rates #1084
+52
−70
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This PR pads audio before feature extraction instead of padding the features, this is inline with how the whisper model was trained because reverting the zero-padded Mel spectrogram features back to time domain results in a white noise with moderate amplitude that causes hallucinations and wrong transcriptions
The
distil-large-v3
WER dropped from 26.04 at 83a368e to 14.472This figure can be reproduced by running
benchmarks/yt_commons.py
and switching the batched inference to sequentialas a side effect, it seems that the transcription is also faster
it took 27 minutes to evaluate the dataset vs 35 minutes before the fix
References:
openai/whisper#730 (comment)
openai/whisper#838 (comment)