Whisper-turbo model support

[codeMonkey-shin]

https://github.com/openai/whisper/pull/2361/files

Will this model be supported in the future?

[stevevaius2015]

Definitely we need this.

[andyhoeung]

I know large-v3 causes some weird stuff in transcription compared to large-v2, I wonder if they improved upon that while also making it faster.

[wangfeng35]

+1 need turbo

[mp3pintyo]

+1 need turbo

[AlanHuang99]

+1 need turbo

[LoggeL]

Downloaded the files for turbo manually and replacing an existing model seems to work as a workaround

https://huggingface.co/openai/whisper-large-v3-turbo/tree/main

[andyhoeung]

I didn't have to replace any model. Just downloaded any faster-whisper large-v3 turbo variant, for example this one https://huggingface.co/Infomaniak-AI/faster-whisper-large-v3-turbo, created a folder in _models called 'faster-whisper-large-v3-turbo', and used '--model=large-v3-turbo'.

[juntaosun]

I didn't have to replace any model. Just downloaded any faster-whisper large-v3 turbo variant, for example this one https://huggingface.co/Infomaniak-AI/faster-whisper-large-v3-turbo, created a folder in _models called 'faster-whisper-large-v3-turbo', and used '--model=large-v3-turbo'.

--model=large-v3-turbo

` Warning: 'large-v3' model may produce inferior results, try 'large-v2'!

Traceback (most recent call last): File "D:\whisper-fast_XXL_main_.py", line 1668, in File "D:\whisper-fast_XXL_main_.py", line 1595, in cli File "faster_whisper\transcribe.py", line 1456, in restore_speech_timestamps File "faster_whisper\transcribe.py", line 798, in generate_segments File "faster_whisper\transcribe.py", line 1109, in encode ValueError: Invalid input features shape: expected an input with shape (1, 128, 3000), but got an input with shape (1, 80, 3000) instead [15868] Failed to execute script 'main' due to unhandled exception! `

faster-whisper-xxl.exe

[ZeVince]

I didn't have to replace any model. Just downloaded any faster-whisper large-v3 turbo variant, for example this one https://huggingface.co/Infomaniak-AI/faster-whisper-large-v3-turbo, created a folder in _models called 'faster-whisper-large-v3-turbo', and used '--model=large-v3-turbo'.

--model=large-v3-turbo

` Warning: 'large-v3' model may produce inferior results, try 'large-v2'!

Traceback (most recent call last): File "D:\whisper-fast_XXL__main__.py", line 1668, in File "D:\whisper-fast_XXL__main__.py", line 1595, in cli File "faster_whisper\transcribe.py", line 1456, in restore_speech_timestamps File "faster_whisper\transcribe.py", line 798, in generate_segments File "faster_whisper\transcribe.py", line 1109, in encode ValueError: Invalid input features shape: expected an input with shape (1, 128, 3000), but got an input with shape (1, 80, 3000) instead [15868] Failed to execute script 'main' due to unhandled exception! `

faster-whisper-xxl.exe

Same error here... any fix since then ?

[Purfview]

Same error here... any fix since then ?

Not yet, but will be sooner than later, just that now some other things have priority.

[Purfview]

Will this model be supported in the future?

It was always supported, as any other custom finetuned model.

Autodownload for it is added in v193.1

[nebehr]

Is turbo model supposed to do any translation at all? Produces untranslated German text with --task translate, whereas vanilla large-v3 appears to work fine.

[Purfview]

"Whisper turbo was fine-tuned for two more epochs over the same amount of multilingual transcription data used for training large-v3, i.e. excluding translation data, on which we don’t expect turbo to perform well."

[nebehr]

I see. From this description though I would expect it to be bad translation, not no translation at all. Anyway, this is beyond the scope of this project.