Evaluation of Deep Learning Approaches to Text-to-Speech Systems for European Portuguese

Abstract

Deep Learning models are considered state-of-the-art regarding Text-to-Speech, displaying very natural and realistic results. However, it is known that these machine learning methods usually require large amounts of data to operate properly. Due to this, an assessment of the system’s ability to generalize to different instances becomes relevant, specially when learning from small data sets to create new voices. This study describes the assessment of a deep learning approach to TTS for European Portuguese. We show that we can use transfer learning techniques to fine-tune a Tacotron-2 model to a specific voice, while preserving speaker identity, without requiring large amounts of data. We also perform a comparison between the developed model and a statistical parametric speech synthesizer enhanced by deep learning, concluding that Tacotron-2 provided an overall better word pronunciation, naturalness and intonation.