Non-Autoregressive Fully Parallel Deep Convolutional Neural Speech Synthesis

Abstract

Deep learning-based speech synthesis evolves by employing a sequence-to-sequence (seq2seq) structure with an attention mechanism. The seq2seq speech synthesis model consists of a pair of the encoder for delivering the linguistic features and the decoder for predicting the mel-spectrogram, and learns the alignment between text and speech through the attention mechanism. The decoder predicts the mel-spectrogram by an autoregressive flow that considers the current input and what they have learned from previous inputs. This is beneficial when processing the sequential data, as in speech synthesis. However, the recursive generation of speech typically requires extensive training time, which slows the speed of synthesis. To overcome these obstacles, we propose a non-autoregressive framework for fully parallel deep convolutional neural speech synthesis. Firstly, we design a new synthesis paradigm that integrates a time-varying metatemplate (TVMT), whose length is modeled with a separate conditional distribution, to prepare the decoder input. The decoding step converts the TVMT into spectral features, which eliminates the autoregressive flow. Secondly, we propose a structure that uses multiple decoders interconnected by up-down chains with an iterative attention mechanism. The decoder chains distribute the burden of decoding, progressively infusing the information obtained from the training target example into the chains to refine the predicted spectral features at each decoding step. For each decoder, the attention mechanism is repeatedly applied to produce the elaborated alignment between the linguistic features and the TVMT, which is gradually transformed into the spectral features. The proposed architecture substantially improves the synthesis speed, and the resulting speech quality is superior to that of a conventional autoregressive model.