SeamlessM4T support in ASR

[nikochiko]

Branched from #168

Removed TTS support because in TTS,:

• non-english TTS was being cut short sometimes
• english TTS was being suddenly broken at 20 seconds

Q/A checklist

• Do a code review of the changes
• Add any new dependencies to poetry & export to requirementst.txt (poetry export -o requirements.txt)
• Carefully think about the stuff that might break because of this change
• The relevant pages still run when you press submit
• If you added new settings / knobs, the values get saved if you save it on the UI
• The API for those pages still work (API tab)
• The public API interface doesn't change if you didn't want it to (check API tab > docs page)
• Do your UI changes (if applicable) look acceptable on mobile?

[SanderGi]

Regarding TTS support, I believe the reason it cuts short and breaks at longer files is that our chunking does not use a stride (see Asana task regarding "Better chunking (with overlap), see hugging face - https://huggingface.co/blog/asr-chunking"). This also applies to ASR and the other SeamlessM4T functionality but is less obvious to observe in this case. I think it would be better to fix the gooey-gpu side chunking rather than remove TTS. Here's one of the ASR outputs I got due to the chunking being off (running on Hugging Face with correct chunking works): "the bone the bone the bone the bone the bone the bone the bone the bone the bone the bone the bone the bone the bone the bone the bone the bone the bone the bone the bone the bone the bone the bone the bone the bone the bone the bone the bone the bone the bone the bone the bone the bone the bone the bone the bone" lol

[devxpy]

running on Hugging Face with correct chunking works

How did you run it on huggingface?

[devxpy]

Oh, transformers supports seamlessm4t now - https://huggingface.co/facebook/hf-seamless-m4t-large

[devxpy]

We should be able to repurpose the whisper code? https://github.com/dara-network/gooey-gpu/blob/3cf4d4393ca719410ae19820e36bd829fab5243c/common/whisper.py#L16

[nikochiko]

@devxpy i looked into it, the seamless code isn't merged into hugging face yet -- https://github.com/huggingface/transformers/pull/25693/files.

Would it be okay to use this HF branch to pull the package from until it is released?