-
-
Notifications
You must be signed in to change notification settings - Fork 1.1k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. Weโll occasionally send you account related emails.
Already on GitHub? Sign in to your account
AssertionError: non-negative timestamp expected #85
Comments
1.zip
And it still outputs these vtt & txt files which seem broken (too short, misaligned, missing the whisper outputs, etc): |
I am pretty sure the reason why this still happens here is because, instead of being based on the official repo of whisper, this project holds copies of outdated code from the offical Whisper's project. Seems like this problem has been solved as described here: #810 and here: #914 You can take a look at the fix here: Fix infinite loop caused by incorrect timestamp tokens prediction Which is not presents in WhisperX's code -> whisperx/decoding.py Code from WhisperX should be refactored in order to follow the original code base to avoid this type of problems. |
I managed to move past this issue for this file using the suggestion here: #84 I will verify the original file can complete after I see where #84 lands regarding |
Getting the same assertion error. Command: Align Model: https://huggingface.co/w11wo/wav2vec2-xls-r-300m-korean Audio Clip: https://www.youtube.com/watch?v=PcysuLjtTeo
EDIT: When I was testing other audio clips I realized that the behavior differs depending on the aligh model I am using. With the same audio clip and command linked above, I changed the model to use wav2vec2-large-xls-r-1b-korean-sample5 instead and it ran without any errors. The same behavior of one model working but another not occurs again for another audio clip of a song. Found another example of one model working and another resulting in "non-negative timestamp expected" |
whisper fix and VAD filtering means this cant happen any more in theory :') |
I'm not sure how a negative timestamp could have been generated, but I seem to have done it ๐ .
When I run
whisperx "/mnt/d/Editing/Audiobooks/a7/a7.wav" --language Japanese --output_dir "/mnt/d/Editing/Audiobooks/a7/" --model large-v2 --vad_filter --align_model WAV2VEC2_ASR_LARGE_LV60K_960H --hf_token some_token
I get
I'm happy to share the 6 hour file for testing purposes on request. I tried breaking the audio into a small clip, and tested on #84, but unfortunately ended up with other errors.
The text was updated successfully, but these errors were encountered: