As someone who thinker with Text to Speech (TTS), I can say this apply to TTS as well. Good model such as Tacotron2 rarely scale beyond clean ( good text and speech alignmen ) large ( > 12 hours ) datasets.