Empirical Evaluation of Deep Learning Model Compression Techniques on the WaveNet Vocoder

TL;DR

This paper evaluates various model compression techniques to accelerate WaveNet vocoders, achieving significant size reduction without sacrificing audio quality, facilitating deployment in scalable text-to-speech systems.

Contribution

It provides a comprehensive empirical comparison of sparsity and quantization methods for WaveNet, demonstrating effective compression ratios with maintained fidelity.

Findings

Achieved up to 13.84x compression without quality loss

Compared multiple sparsity and quantization techniques

Implemented using open source frameworks for wider adoption

Abstract

WaveNet is a state-of-the-art text-to-speech vocoder that remains challenging to deploy due to its autoregressive loop. In this work we focus on ways to accelerate the original WaveNet architecture directly, as opposed to modifying the architecture, such that the model can be deployed as part of a scalable text-to-speech system. We survey a wide variety of model compression techniques that are amenable to deployment on a range of hardware platforms. In particular, we compare different model sparsity methods and levels, and seven widely used precisions as targets for quantization; and are able to achieve models with a compression ratio of up to 13.84 without loss in audio fidelity compared to a dense, single-precision floating-point baseline. All techniques are implemented using existing open source deep learning frameworks and libraries to encourage their wider adoption.