I am attempting to fine-tune ‘whisper_tiny’ on a French dataset. Initially, the Word Error Rate (WER) metric and the evaluation loss show a decrease. However, after a certain point, both metrics start to increase. I attempted to enhance the quality of transcriptions by removing special characters. This approach was taken due to the nature of the French language, which is characterized by a wide variety of characters. Unfortunately, despite these efforts, no noticeable improvements have been observed. I am uncertain about the reasons behind this outcome.
this is the link of the notebook : notebook link
this are training and evaluation loss and wer values during training: