Improvement of Minimum Statistical Spectrum for Beamforming

In this paper, we propose a method to improve the noise suppression performance of adaptive beamforming in environments where the speech source moves or where the noise level is very high. Among various speech enhancement methods, adaptive beamforming based on microphone array has high noise suppression performance. However, the performance of adaptive beamforming degrades in highly noisy environment and it cannot track the speech source when the speaker moves. To solve the problem, the proposed method utilizes the characteristics that the amplitude spectrum of noise varies more significantly than that of the speech in a short period. In addition to spectra obtained from each input speech signal of microphone arrays, spectra obtained at previous frames of each input speech signal are effectively utilized to find the Minimum Statistical Spectrum (MSS). From the resulting MSS, an enhanced speech signal is created and used in adaptive beamforming with the input speech signals of microphone array. Experiments show that the proposed method provides a high performance relative to the original MSS approach.