Filtering the training data

[ANonEntity]

Whisper tends to transcribe/translate silence as "Thank you for watching!", "Please subscribe to my channel!" and so on, since the training data contains YouTube captions. It seems like removing these lines from the training data, and then retraining/finetuning Whisper is the only way to solve this.

Would it be possible to detect unvoiced lines like these automatically? Maybe by filtering all the training data through a VAD?