Using Intermicrophone Correlation to Detect Speech in Spatially Separated Noise

Abstract

This paper describes a system for determining intervals of "high" and "low" signal-to-noise ratios when the desired signal and interfering noise arise from distinct spatial regions. The correlation coefficient between two microphone signals serves as the decision variable in a hypothesis test. The system has three parameters: center frequency and bandwidth of the bandpass filter that prefilters the microphone signals, and threshold for the decision variable. Conditional probability density functions of the intermicrophone correlation coefficient are derived for a simple signal scenario. This theoretical analysis provides insight into optimal selection of system parameters. Results of simulations using white Gaussian noise sources are in close agreement with the theoretical results. Results of more realistic simulations using speech sources follow the same general trends and illustrate the performance achievable in practical situations. The system is suitable for use with two microphones in mild-to-moderate reverberation as a component of noise-reduction algorithms that require detecting intervals when a desired signal is weak or absent.