LightSpeech: Lightweight and Fast Text to Speech with Neural Architecture Search

Authors

• Renqian Luo (University of Science and Technology of China) lrq@mail.ustc.edu.cn
• Xu Tan (Microsoft Research) xuta@microsoft.com
• Rui Wang (Microsoft Research) ruiwa@microsoft.com
• Tao Qin (Microsoft Research) taoqin@microsoft.com
• Enhong Chen (University of Science and Technology of China) cheneh@ustc.edu.cn
• Tie-Yan Liu (Microsoft Research) tyliu@microsoft.com

Abstract

Text to speech (TTS) has been broadly used to synthesize natural and intelligible speech in different scenarios. Deploying TTS in various end devices such as mobile phones or embedded devices requires extremely small memory usage and inference latency. While non-autoregressive TTS models such as FastSpeech have achieved significantly faster inference speed than autoregressive models, their model size and inference latency are still large for the deployment in resource constrained devices. In this paper, we propose LightSpeech, which leverages neural architecture search (NAS) to automatically design more lightweight and efficient models based on FastSpeech. We first profile the components of current FastSpeech model and carefully design a novel search space containing various lightweight and potentially effective architectures. Then NAS is utilized to automatically discover well performing architectures within the search space. Experiments show that the model discovered by our method achieves 15x model compression ratio and 6.5x inference speedup on CPU with on par voice quality.

Audio Samples

All of the audio samples use Parallel WaveGAN (PWG) as vocoder.

Compare to FastSpeech 2

*The general solidity of a page is much to be sought for.

*The modern printer, in the teeth of the evidence given by his own eyes, considers the single page as the unit, and prints the page in the middle of his paper.

*For where these are boldly and carefully designed, and each letter is thoroughly individual in form.

*The words may be set much closer together, without loss of clearness.

*No definite rules, however, except the avoidance of “rivers” and excess of white, can be given for the spacing.

Ablation Study

FastSpeech 2*: manually designed lightweight FastSpeech 2 model.

*Which requires the constant exercise of judgment and taste on the part of the printer.

*Only nominally so, however, in many cases, since when he uses a headline he counts that in.

*From the time when books first took their present shape till the end of the sixteenth century, or indeed later.

*The paper on which the printing is to be done is a necessary part of our subject.

*Of this it may be said that though there is some good paper made now.

Our Related Works

Neural Architecture Optimization
Neural Architecture Search with GBDT
FastSpeech: Fast, Robust and Controllable Text to Speech
FastSpeech 2: Fast and High-Quality End-to-End Text to Speech
LRSpeech: Extremely Low-Resource Speech Synthesis and Recognition
Almost Unsupervised Text to Speech and Automatic Speech Recognition