Low distortion speech enhancement in the DCT domain using optimal estimate of the a priori SNR

Abstract

The spectral subtraction based algorithms are commonly used for single channel speech enhancement because of their elegant performance in denoising with low cOlllputationalload. They, however, suffer from a serious drawback in that the enhanced speech is accompanied by unpleasant musical noise artifact, which is characterized by tones with random frequencies. It is known that the key point behind the reduction of musical noise by the minimum-mean-squared-error (MMSE) estimator is the use of a priori SNR. The "decision-directed" approach widely used for its estimation requires an averaging parameter. Conventionally, a constant value is chosen by most researchers. The main objective of this work is the development of a self-adaptive smoothing parameter in the MMSE sense to estimate the a priori SNR in the DCT domain which can account for the abrupt changes in the speech spectral amplitudes. The performance improvement using the proposed self-adaptive smoothing parameter in the commonly used spectral subtraction algorithms for denoising speech corrupted by background noise is noteworthy. The conventional Wiener filtering shows better denoising performance in terms of overall and average segmental SNRs with the cost paid in Itakura-Saito (IS) measure as compared to the spectral subtraction based methods. In this work, a generalized Wiener filter is proposed to improve the IS measure without sacrificing enhanced speech quality in terms of SNR by introducing a new term in the gain function. A comparative study with the spectral subtraction algorithms and the conventional Wiener filter confirms the superiority of the proposed generalized Wiener filter.