Persian ASR with Whisper

[Mohammad Reza Taesiri]

Hello All,

This is my first post here; nice to "e-meet" you all. A while ago, I fine-tuned the whisper model using the Farsi subset of the Mozilla Common Voice dataset. You can try the (best) model here: https://huggingface.co/taesiri/whisper-small-fa-7

The training loss is almost zero, but the final WER is around 34. This means further training without more data (or some magical tricks) won't improve the performance. 

Best,

Mohammad Reza

[Edward Pearce]

Dear Mohammad Reza, are you still interested in the topic? if yes, can we have a little chat? I might have the data to push the WER further down. 

[Mohammad Reza Taesiri]

Hello Edward, Yes, for sure! I have the computation to train another model, just let me know.