Spaces:
Runtime error
Apply for community grant: Personal project
Hello,
@sanchit-gandhi
and
@reach-vb
!
I would like to apply for community grant to launch this demo of a Belarusian Whisper Small model on a GPU for faster inference.
Training is still in progress. The current WER on a Belarusian Common Voice 11 validation split is 6.464%. I have not evluated model on a test split of the same dataset because the training is not finished yet.
Original openai/whisper-small model was not evaluated on a Common Voice dataset according to a section D.2.2 of a paper https://cdn.openai.com/papers/whisper.pdf. It was evaluated on a Fleurs (section D.2.4) - Belarusian Whisper Small had 75.1% WER reported.
Once training will finish, I will evaluate my model both on test set of Common Voice 11 and on Fleurs as well
First training run have finished - validation WER on Common Voice 11 is 6.3672%
I'm planning to laungh a second run tomorrow and will try to decrease WER even further.
I've evaluated model on a test set of Common Voice 11. test WER = 6.79%