Comment by blackcat201

Do you get good results when adding scraped youtube audio? My model performance on LibriSpeech dev drops a bit when adding youtube audio to the training dataset ( my guess is likely due to poor alignment from auto generated captions ).