Model-Based Post Filter for Microphone Array Speech Enhancement

Generalized sidelobe canceller (GSC) is wildly used in speech enhancement due to its efficient implementation. However, the conventional GSC has some drawbacks when applied to speech enhancement system. First, it is focused on improving the signal-to-noise ratio (SNR) without considering the characteristics of speech so that is not optimal for speech enhancement applications. Second, the adaptive branch in the GSC does not always estimate the noise in the fixed branch output accurately, especially when the SNR is high, the noise is spatially incoherent, or the spatial incoherent noises and spatial coherent interferences coexist. In this paper, we propose a model-based post filter for the sub-band GSC which is a typical form of the microphone array beamformer. An improved noise estimation method is developed to estimate the noise in the fixed branch output of each sub-band GSC from its adaptive branch output. Then the fixed branch output is filtered by an optimal filter which is constructed according to a GMM model trained by clean speeches and an online-estimated noise model. Experimental results show that the proposed method achieves significant improvement over the conventional sub-band GSC and outperforms several speech enhancement methods in different noisy environments.

I. Introduction

Environmental noises and interferences usually cause significant performance degradations in many speech applications such as speech communication, hearing aids, and automatic speech recognition. Therefore speech enhancement, which aims to suppress the environmental noise without distorting the target speech, is required to improve the speech quality, intelligibility or the performance of speech coding and recognition systems [1].