Hindi ASR: Fine-Tuning Wav2Vec2

Hi guys! I am planning to work on fine-tuning XLSR-Wav2Vec2 in Hindi. Would love to collaborate with others working on it too. We can share and discuss issues here. :slight_smile:

3 Likes

hii @hadarishav i am also working on it tooo. :star_struck:

1 Like

Hi @danurahul, @hadarishav

I don’t think common voice has ample amount of Hindi data, the given data when trained doesn’t lead to any reduction in the WER.

Let me know if there are any other datasets that you find :slight_smile:

Thanks,
Gunjan

https://drive.google.com/drive/folders/1s5bAozkso28BLCvwG0jE2T0Oj_ONEejv

check this data its very huge

1 Like

Hi @danurahul,

Thanks a lot, I’ll look into it.

Do you have a LICENSE to use this data? or a paper you can cite?

Its freely availablr for competition

Joining a little late to the party but would love to collaborate too for Hindi ASR.

Has anyone fine-tuned on a model on Hindi yet?

1 Like

@danurahul Thanks for sharing the resource. This looks promising.
@gchhablani You can find more details on the dataset here - GitHub - Speech-Lab-IITM/Hindi-ASR-Challenge: 🎯 Speech Recognition Challenge by Speech Lab - IIT Madras
They make the dataset freely available for research. So I think, we can use it safely!
I have started working today. Reading the docs and other instructions. Will get to the code soon. Will start with exploring the data first. How about you?

1 Like

Does anyone know of a code-snippet to load the dataset from a drive for use in this fine-tuning task? Much appreciated!!

Hello :wave:,

I am working on Hindi as well. Currently I am able to achieve 0.54 validation WER as well as 72 percent WER on test dataset. I will be updating the full info soon.

@theainerd Which dataset are you using?

I am using Open SLR dataset. I tried training on common voice but the dataset is too small and test WER remain 1 only.

1 Like

Does anyone have access to this database:

https://www.iitm.ac.in/donlab/tts/database.php

It seems they do not give access unless you are affiliated to an institute.

Hi,

Look up mounting drive on Colab :slight_smile:
If you’re using a local setup, then I’m not fully sure.

Which Open SLR dataset? Can you share the link here please? Is it the INTERSPEECH 2021 one?

Hi, joined late… Would like to support others working on finetuning hindi ASR
-Sagar

hey, I’m working in the ASR domain. let me know how to get in touch so we can collaborate!