Augmenting datasets for training audio generation models

Abstract

A target voice dataset may be augmented using speech prediction. Encoder and decoder models may be trained to encode audio data into encoded speech data, and convert it back to audio. The encoded units may include semantic information (e.g., phonemes and/or words) as well as feature data indicating prosody, timbre, speaker identity, speech style, emotion, etc. of speech. An acoustic/semantic language model (ASLM) may be configured to predict encoded speech data in a manner analogous to a language model predicting words; for example, based on preceding encoded speech data. The models may be used to generate synthesized speech samples having voice characteristics (e.g., feature data) similar to those of the target voice dataset. The augmented dataset may be used to train a text-to-speech (TTS) model to reproduce the target voice characteristics, and may improve performance of the TTS model over training with only the original target voice dataset.