Residual Noise Removal on Beamforming for Robust Hands-free Speech Recognition

Abstract

A simple and efficient design for residual noise removal is proposed with spectral subtraction and microphone array processing for robust speech recognition. The overall system is a three-stage approach. The system first constructs the delay-and-sum beamformer, and then applies noise spectral subtraction and residual noise removal in the following post-processing stages. The residual noise spectrum of each frame of the enhanced speech signal is estimated by the Wiener filtering technique and removed by spectral subtraction. Experimental results show that the proposed algorithm improves the noisy speech recognition accuracy significantly.