An Utilizing of Spectral Mask to Enhance GSC Beamformer

Abstract

Speech source separation is always a complex problem when the recording situation has interference, unwanted noise, transport vehicle or third-party talker. Without any assumed information about source, it is very difficult to separate useful signal. Nowadays, there has been a several approaches to extract target directional speech. This technology, which use microphone array to utilize the microphone placement and speaker location, named beamformer. Generalized Sidelobe Canceller (GSC) Beamformer is one of the most popular method for remov- ing interfering noise and achieve desired signal with high speech quality. However, its performance is always degraded due to many reasons, such as the different microphone sensitivities, microphone mismatches, the error of the direction of arrival (DOA). In this paper, the author presented a spectral mask, which uses the instantaneous value of the signal-to-noise SNR for enhancing GSC's evaluation in presence of complicated noise. The developed spectral mask is based on the assumption of the amount of speech leakage is proportional with the ratio SNR, the higher SNR, the smaller spectral mask. The performed experiment was verified and confirmed of this approach through the numerical results of the SNR from 2.0 (dB) to 6.9 (dB). The proposed method, which provides us an efficient solution to deal realistic speech task with microphone array, can be integrated into multi-microphone system with arbitrary geometry or distribution of microphones.