Non-Autoregressive TTS with Explicit Duration Modelling for Low-Resource Highly Expressive Speech

• DOI: 10.21437/ssw.2021-17
• Publication Date: 2021-08-24T06:58:52Z
• Authors: Raahil Shah; Kamil Pokora; Abdelhamid Ezzerg; Viacheslav Klimkov; Goeric Huybrechts; Bartosz Putrycz; Daniel Korzekwa; Thomas Merritt
• Keywords: Similarity (geometry); Generative model

Whilst recent neural text-to-speech (TTS) approaches produce high-quality speech, they typically require a large amount of recordings from the target speaker. In previous work, 3-step method was proposed to generate TTS while greatly reducing data required for training. However, we have observed ceiling effect in level naturalness achievable highly expressive voices when using this approach. paper, present building with as little 15 minutes speech Compared current state-of-the-art approach, our improvements close gap by 23.3% and 16.3% speaker similarity. Further, match similarity Tacotron2-based full-data (~10 hours) model only data, whereas 30 or more, significantly outperform it. The following are proposed: 1) changing an autoregressive, attention-based non-autoregressive replacing attention external duration 2) additional Conditional Generative Adversarial Network (cGAN) based fine-tuning step.