The Modified Post-Filter Beamforming for Speech Enhancement

A modified Wiener post-filter beamforming for speech enhancement is presented. The proposed algorithm focuses on the estimation of the ideal signals power spectrum for the Wiener filter, both auto-correlation and cross-correlation are taken into consideration to obtain as more power spectrum estimations as possible, the average of which produces more accurate result. Meanwhile, the alteration of cross-correlation caused by the moving speaker is revised. The performance of the algorithm has been evaluated from Signal to Noise Ratio (SNR), Average Log Spectrum Distance (ALSD) and speech recognition rate. The SNR is improved by 5 dB. The command recognition rate based on Hidden Markov Model (HMM) in the car environment reaches 84%. The low frequency noise that still remains after the process of the primary method is reduced by the proposed method, and the speech signal distortion is also decreased.