End-to-end speech enhancement based on discrete cosine transform

TL;DR

This paper introduces a novel speech enhancement method using Discrete Cosine Transform (DCT) within a U-net framework to reconstruct valid short-time spectra, achieving superior performance over traditional methods.

Contribution

It proposes a new DCT-based approach for speech enhancement that ensures a valid short-time spectrum, improving upon prior magnitude-only estimation methods.

Findings

Achieves better speech quality than magnitude-only methods.

Uses DCT to reconstruct valid short-time spectra.

Employs U-net architecture for enhancement.

Abstract

Previous speech enhancement methods focus on estimating the short-time spectrum of speech signals due to its short-term stability. However, these methods often only estimate the clean magnitude spectrum and reuse the noisy phase when resynthesize speech signals, which is unlikely a valid short-time Fourier transform (STFT). Recently, DNN based speech enhancement methods mainly joint estimation of the magnitude and phase spectrum. These methods usually give better performance than magnitude spectrum estimation but need much larger computation and memory overhead. In this paper, we propose using the Discrete Cosine Transform (DCT) to reconstruct a valid short-time spectrum. Under the U-net structure, we enhance the real spectrogram and finally achieve perfect performance.