Add a task for automatic text recognition

[PonteIneptique]

Hi :)
We are in the process of working a pipeline to help people publish their data to huggingface in the context of HTR/OCR groundtruth and HTR-United, and have ourselves a fair amount of data.
I wonder if it could be possible to have a ATR (Automatic Text Recognition) or OCR/HTR (Optical Character Recognition / Handwritten Text Recognition) task to register our datasets under, instead of the quite broader Vision to Text, which seems more focused on image-description datasets ?
Thanks !

[coyotte508]

cc @merveenoyan @osanseviero

[osanseviero]

cc @sanchit-gandhi and @Vaibhavs10 for our audio experts :)

[Vaibhavs10]

This is more vision no?

[PonteIneptique]

This is more Vision than this is Text (although, depending and who you ask...) but I don't think that Multimodal > Vision-to-text is a good match for HTR/OCR/ATR

[osanseviero]

Sorry for my confusion, I read too quickly and did string matching with ASR 🥲

Yes, this is indeed vision, In the past, OCR models have been tagged as image-to-text such as in https://huggingface.co/microsoft/trocr-base-handwritten . I think potentially we could keep image-to-text + add a secondary subtype for this use case (either ocr or atr as suggested). WDYT @merveenoyan @NielsRogge @lhoestq ?

[lhoestq]

I'm ok to add a new task_id "ocr" or "optical-character-recognition" under "image-to-text"

[merveenoyan]

I agree with @lhoestq.