Deep Denoising Auto-encoder for Statistical Speech Synthesis

TL;DR

This paper introduces a deep denoising auto-encoder to improve acoustic feature extraction for speech synthesis, resulting in higher quality synthetic speech through a non-linear, data-driven approach.

Contribution

It presents a novel deep denoising auto-encoder method for extracting acoustic features, outperforming traditional mel-cepstral analysis in speech synthesis.

Findings

Enhanced speech quality in analysis-by-synthesis experiments

Improved naturalness in text-to-speech synthesis

Effective non-linear feature extraction demonstrated

Abstract

This paper proposes a deep denoising auto-encoder technique to extract better acoustic features for speech synthesis. The technique allows us to automatically extract low-dimensional features from high dimensional spectral features in a non-linear, data-driven, unsupervised way. We compared the new stochastic feature extractor with conventional mel-cepstral analysis in analysis-by-synthesis and text-to-speech experiments. Our results confirm that the proposed method increases the quality of synthetic speech in both experiments.