FastSpeech: Fast, Robust and Controllable Text to Speech

TL;DR

FastSpeech tackles slow, fragile autoregressive TTS by introducing a non-autoregressive, parallel mel-spectrogram generator based on a Feed-Forward Transformer. It employs a length regulator and a duration predictor trained with teacher alignments and sequence-level distillation to align phonemes with mel frames and enable speed and prosody control. On LJSpeech, it achieves near-parity with autoregressive models in quality while delivering 270x speedup for mel-spectrogram and 38x end-to-end synthesis, plus improved robustness and flexible length control. This work enables practical, fast, robust, and controllable TTS with potential extensions to multi-speaker and low-resource scenarios.

Abstract

Neural network based end-to-end text to speech (TTS) has significantly improved the quality of synthesized speech. Prominent methods (e.g., Tacotron 2) usually first generate mel-spectrogram from text, and then synthesize speech from the mel-spectrogram using vocoder such as WaveNet. Compared with traditional concatenative and statistical parametric approaches, neural network based end-to-end models suffer from slow inference speed, and the synthesized speech is usually not robust (i.e., some words are skipped or repeated) and lack of controllability (voice speed or prosody control). In this work, we propose a novel feed-forward network based on Transformer to generate mel-spectrogram in parallel for TTS. Specifically, we extract attention alignments from an encoder-decoder based teacher model for phoneme duration prediction, which is used by a length regulator to expand the source phoneme sequence to match the length of the target mel-spectrogram sequence for parallel mel-spectrogram generation. Experiments on the LJSpeech dataset show that our parallel model matches autoregressive models in terms of speech quality, nearly eliminates the problem of word skipping and repeating in particularly hard cases, and can adjust voice speed smoothly. Most importantly, compared with autoregressive Transformer TTS, our model speeds up mel-spectrogram generation by 270x and the end-to-end speech synthesis by 38x. Therefore, we call our model FastSpeech.

Figures (4)

• Figure 1: The overall architecture for FastSpeech. (a). The feed-forward Transformer. (b). The feed-forward Transformer block. (c). The length regulator. (d). The duration predictor. MSE loss denotes the loss between predicted and extracted duration, which only exists in the training process.
• Figure 2: Inference time (second) vs. mel-spectrogram length for FastSpeech and Transformer TTS.
• Figure 3: The mel-spectrograms of the voice with 1.5x, 1.0x and 0.5x speed respectively. The input text is "For a while the preacher addresses himself to the congregation at large, who listen attentively".
• Figure 4: The mel-spectrograms before and after adding breaks between words. The corresponding text is "that he appeared to feel deeply the force of the reverend gentleman's observations, especially when the chaplain spoke of". We add breaks after the words "deeply" and "especially" to improve the prosody. The red boxes in Figure \ref{['fig_add_break']} correspond to the added breaks.