Persian ASR: Fine-Tuning Wav2Vec2

Hi Persian speakers,

Let’s do something cool for Persian together. Let me start first, I fine-tuned the model on only 40% of the whole dataset and got a 0.5 WER score. The preprocessing step considers only 36 characters in Farsi, which I think is covered all.

After a few more epochs and training the whole data, I got 32.09% WER and 8.23% CER on the test set.

Give it a try! I provided the preprocessing and the training arguments in the following link.