Datasets:
Transcripts misaligned / wrong
#7
by RonanMcGovern - opened
Quite a few transcripts appear wrong, based on test set inspection. The audio doesn't align to the text.
Can you please give me a list of utterances wrongly aligned? Thanks a lot!
EDIT: My previous analysis of Whisper v3 CER on the dataset was flawed as a number of the transcription attempts failed and gave falsely high CER.
Here is the updated histogram:
and I've put a dataset here of the rows that have high CER. It's only about 206 rows out of a total of 4683 (note that the total rows don't correspond exactly to your dataset because I already remove some rows and cut out certain text/audio within rows based on alignment confidence).