Fine-tuning code
Are you able to share your fine-tuning code? Your results are really impressive!
Thanks! I actually don't know anything about the current state of ASR for Frisian: I just picked it as a nice example of a language that is not included in the Whisper training set and that I - as a Dutch person - am somewhat able to understand.
I could share a notebook with my training code, but it would basically be a one-to-copy of this public Huggingface draft blog post (readable form). The only thing I had to change is this bugfix and of course change the language of the dataset from dv to fy-NL.
I will make my notebook public after they publish their blog post. It trains in a couple of hours on a Colab Pro "premium" GPU.
Their blog post is out! Fine-Tune Whisper For Multilingual ASR with 🤗 Transformers.
They've made some fixes to the Whisper implementation in Huggingface Transformers that make my notebook obsolete, so just do whatever they do.