Research Area:  Machine Learning
Neural sequence-to-sequence text-to-speech synthesis (TTS) can produce high-quality speech directly from text or simple linguistic features such as phonemes. Unlike traditional pipeline TTS, the neural sequence-to-sequence TTS does not require manually annotated and complicated linguistic features such as part-of-speech tags and syntactic structures for system training. However, it must be carefully designed and well optimized so that it can implicitly extract useful linguistic features from the input features. In this paper we investigate under what conditions the neural sequence-to-sequence TTS can work well in Japanese and English along with comparisons with deep neural network (DNN) based pipeline TTS systems. Unlike past comparative studies, the pipeline systems also use neural autoregressive (AR) probabilistic modeling and a neural vocoder in the same way as the sequence-to-sequence systems do for a fair and deep analysis in this paper. We investigated systems from three aspects: a) model architecture, b) model parameter size, and c) language. For the model architecture aspect, we adopt modified Tacotron systems that we previously proposed and their variants using an encoder from Tacotron or Tacotron2. For the model parameter size aspect, we investigate two model parameter sizes. For the language aspect, we conduct listening tests in both Japanese and English to see if our findings can be generalized across languages. Our experiments on Japanese demonstrated that the Tacotron TTS systems with increased parameter size and input of phonemes and accentual type labels outperformed the DNN-based pipeline systems using the complicated linguistic features and that its encoder could learn to compensate for a lack of rich linguistic features. Our experiments on English demonstrated that, when using a suitable encoder, the Tacotron TTS system with characters as input can disambiguate pronunciations and produce natural speech as good as those of the systems using phonemes. However, we also found that the encoder could not learn English stressed syllables from characters perfectly and hence resulted in flatter fundamental frequency. In summary, these experimental results suggest that a) a neural sequence-to-sequence TTS system should have a sufficient number of model parameters to produce high quality speech, b) it should also use a powerful encoder when it takes characters as inputs, and c) the encoder still has a room for improvement and needs to have an improved architecture to learn supra-segmental features more appropriately.
Author(s) Name:  Yusuke Yasuda,Xin Wang,Junichi Yamagishi
Journal name:  Computer Speech & Language
Publisher name:  Elsevier
Volume Information:   Volume 67, May 2021, 101183
Paper Link:   https://www.sciencedirect.com/science/article/abs/pii/S0885230820301169