How do you transcribe a long audio file?

[nickponline]

I have a 30 min mp3 I would line to transcribe with whisper is there a command line that will do this, or chunk it up automatically?

[abhinavkulkarni]

@nickponline: Chunking is done automatically.

[jongwook]

@nickponline We're thinking of supporting a callback or making a generator version of transcribe() (some discussions in #1025).

@masafumimori The OP was about using this Python package and model locally, and the 25MiB limit is a temporary restriction on the maximum file size when using the Whisper API. While we'd like to increase the limit in the future (or better, make an async API), you could split the MP3 in the meantime, or convert to a lower-bitrate MP3 or M4A to make it under 25MiB and submit it to the API for now.

[masafumimori]

@jongwook
I see! I tried that and it worked! Thank you :)

[niuta24]

@masafumimori can you share your code, please? did you do some code to automatically divide audio at 25Mb parts and give them to Whisper?

[masafumimori]

@niuta24
I did something like this.

In my case, it was enough to split into two.
Also, be aware of doing this may split at the middle of sentence being spoken.
Hope this helps!

[SpyCoder77Alt]

is there a way to merge the two transcriptions into one?

[phineas-pta]

you can use ffmpeg to split the audio, for e.g. making 1h segment ffmpeg -hide_banner -i <YOUR AUDIO FILE> -c copy -map 0 -segment_time 1:0:0 -f segment <OUTPUT DIR>/segment_%03d

and then you can loop over the directory with whisper

[omenking]

I would think that this could lead inaccuraries if you are spliting the middle of an utterance, and so I guess you have to have overlap between chunks to get best results?

[SpyCoder77Alt]

yeah