I wonder if that's specifically a result of training it on youtube videos where the audio/subtitles don't actually match (recipe videos with no speech that have additional text added in the subtitles)