Abstract

Close-talk system is used in noisy and reverberant condition. This paper addresses the problem of close talk speech enhancement as a binary classification using dual microphones feature in complex auditory environment. In this work, we investigate a speech segregation framework. Deep learning is employed as a mechanism to find the robustness classifier from the signal difference between the two microphones. A estimated binary mask is used to separate the target speech from the mixture. We systematically use recording and simulated corpus to examine the performance of the proposed algorithm, and generalize to untrained configurations. Three comparison systems are used. Signal to noise ratio (SNR) improvement and automatic speech recognition (ASR) accuracy show that the proposed system achieves robust performance across a variety of noise location, interference type and reverberant condition.

Paper Details