Much difference in evaluation results on google/fleurs dataset
#3
by AICoding91 - opened
Thank you for amazing repo.
I tried to evalute your model on google/fleurs dataset. Because you did not provide evaluation source code, so I based on whisper-small https://huggingface.co/openai/whisper-small#evaluation
I changed dataset google/fleurs, but I got very big WER. It is much different from your reported WER on google/fleurs dataset.
Could you please provide evaluation source code for this model? Thank you so much.
And I also saw, when running inference, you model gives spaces between Japanese characters. It did not happen with the original model whisper-small. Is there any special config in your finetuning process?