Natural TTS Synthesis by Conditioning WaveNet on Mel Spectrogram Predictions

摘要

This paper describes Tacotron 2, a neural network architecture for speechsynthesis directly from text. The system is composed of a recurrentsequence-to-sequence feature prediction network that maps character embeddingsto mel-scale spectrograms, followed by a modified WaveNet model acting as avocoder to synthesize timedomain waveforms from those spectrograms. Our modelachieves a mean opinion score (MOS) of 4.53 comparable to a MOS of 4.58 forprofessionally recorded speech. To validate our design choices, we presentablation studies of key components of our system and evaluate the impact ofusing mel spectrograms as the input to WaveNet instead of linguistic, duration,and F0 features. We further demonstrate that using a compact acousticintermediate representation enables significant simplification of the WaveNetarchitecture.