XLSR-53

[yagan93]

🚀 Feature request

Is it possible to use XLSR-53 with transformers in the near future?

[rodrigoheck]

Apparently, someone just did it. But there are some files missing and it currently unusable. Hopefully the author will soon update it :)

[LysandreJik]

Pinging @patrickvonplaten for knowledge :)

[patrickvonplaten]

Yeah, I just added the pretrained checkpoint. I'll release a notebook by the middle/end of next week on how to fine-tune the checkpoint. Please ping me here again if you can't find it :-)

[yagan93]

@patrickvonplaten Thanks a lot! Cant wait to use it and see how it performs :)

[pasikon]

@patrickvonplaten cool! is it possible to use with Transformers XLSR-53 finetuned with Fairseq?

[patrickvonplaten]

will release a notebook either tomorrow or on Monday about it :-)

[pasikon]

@patrickvonplaten can't wait :)

[patrickvonplaten]

Notebook is available here: https://huggingface.co/blog/fine-tune-xlsr-wav2vec2 :-)

[patrickvonplaten]

We are organizing a "fine-tuning XLSR-53" event. Check this announcement: https://discuss.huggingface.co/t/open-to-the-community-xlsr-wav2vec2-fine-tuning-week-for-low-resource-languages/4467. Would be awesome if you want to participate :-)

[yagan93]

@patrickvonplaten

Hey buddy!

First and foremost I want to thank you again for all your effort! Really appreciate it!

Got another litte question:

Fine tuned a wav2vec-large-xlsr-53 model on Swiss German (bernese dialect) as laid out in one of your blogs.

Currently trying to add an already existing 6-Gram-KenLM on top.

Could you give me some hints on how to do it? Or is it yet not even possible?

Kind regards
Yves 😉

[patrickvonplaten]

Hey Yves,

Here a forum post regarding this issue: https://discuss.huggingface.co/t/language-model-for-wav2vec2-0-decoding/4434

[gretatuckute]

Hi all,
I am following up on this issue: I am trying to use the pre-trained Wav2Vec2-XLSR-53 (https://huggingface.co/facebook/wav2vec2-large-xlsr-53) and according to the documentation, it should be available as:

from transformers import AutoTokenizer, AutoModel
tokenizer = AutoTokenizer.from_pretrained("facebook/wav2vec2-large-xlsr-53")
model = AutoModel.from_pretrained("facebook/wav2vec2-large-xlsr-53")

The model is available, but the tokenizer is not found (error: OSError: Can't load tokenizer for 'facebook/wav2vec2-large-xlsr-53'. Make sure that: (...) ). I tried using Transformers 4.2.2 and 4.5.0 as well as cloning the repository -- no luck. I am able to successfully load e.g. the French version:

tokenizer = AutoTokenizer.from_pretrained("facebook/wav2vec2-large-xlsr-53-french")

But not the base XLSR tokenizer?

Thanks so much for the brilliant work!

[yagan93]

Hey @gretatuckute

Check out my HuggingFace Profile https://huggingface.co/Yves. There you'll find what you're after.
If you ask @patrickvonplaten he could also invite you the wav2vec xlsr slack channel :)

Cheers
Yves

[gretatuckute]

Hi @yagan93, thank you for getting back! On your HF profile I only see the Swiss-German tokenizer?

[yagan93]

@gretatuckute You just got to swop the models and make little adjustments. Check out this notebook for details information on how to do so.

[LysandreJik]

Closed by #10648