Text-to-speech processing

Abstract

A speech-processing system may be configured to generate expressive synthesized speech. The system may include a prosody prediction model that generates a combination of durations and acoustic representations that may be based on the content of the text as well as additional context information. The model may be trained to predict a joint probability between linguistic representations (e.g., derived from text) and combined duration/acoustic representations (e.g., derived from audio). At inference, the model can process linguistic representations derived from text to predict combined duration/acoustic representations. In some implementations, the model may process additional information; for example, semantic embeddings output by a language model based on the text. In another example, the model may receive a speaker embedding representing voice characteristics of a particular speaker. A decoder may process the durations and acoustic representations output by the model to generate audio data representing the synthesized speech and representing expressive prosodic variation.