Multi speaker text to speech transfer learning

Abstract

Speech synthesis is experiencing a breakthrough as progressive leaps in artificial intelligence have led to a shift from the robotic standard voice to a more human-like voice with emotional inflections across multiple speakers and languages. Tacotron has been used intensively for such text-to-speech syntheses lately. Accordingly, in this thesis, I aim at studying the possibility of performing Multispeaker text-to-speech (TTS) transfer learning with Tacotron 2 in French to overcome the need of having multiple machines one per speaker. That is achieved by finetuning the Tacotron 2 training processor to allow learning the multiple speakers available in our dataset. For that, we use publicly available online French datasets that are already annotated. However, the main challenge that such models face is data efficiency and quality of the speaker audio files as well as speaker variability where each speaker might have a different accent or speaking rate. Despite that our model provided us with adequate results when presented with only a few hours of new speakers from different genders.