Issue with word-level timestamp with WhisperX V2

[shruru]

hi there, somehow I couldn't export srt file with word_timestamp enabled. there is not word-level srt file generated. e.g.

whisperx --verbose True --model large-v2 --language en --hf_token XXXXX --word_timestamps True --output_format srt --output_dir . --vad_filter False Video.mp4

Please advise Thank you~

[m-bain]

hey, so you dont need --word_timestamps True this is for the original whisper model. You could play with these to see how it compares, but currently the code doesn't parse the original whisper timestamps.

And to get word-level you need to switch to either --output_format word.srt or --output_format ass (the latter is with highlighting)

[m-bain]

todo (for self): documentation to explain --word_timestamps for OG whisper parse OG whisper word_timestamps for outputting to word.srt "srt-word" for comparison

[shruru]

it works . thanks! but it should be "srt-word" :P

[Saccarab]

so are OG whisper word_confidence values just buried in whisperX result somewhere or is the parameter not enabled on whisperX requests in the first place?

[pdahiya]

in latest version - the .ass format has been removed. How to get word level timestamped transcript now? The VTT or srt files doesn't have word level timestamps. @m-bain can you please help?

[m-bain]

@pdahiya --highlight_words True

outputs srt with highlighted words

[m-bain]

You can also use the json to generate your own, I will add .ass back at some point

[sanxfxteam]

I get this error when trying to output srt-word, what's the way to do it now? whisperx: error: argument --output_format/-f: invalid choice: 'srt-word' (choose from 'all', 'srt', 'vtt', 'txt', 'tsv', 'json', 'aud')