Algorithms for Alignment of Co-Channel Speech Signals in Adaptive Speech Separation

Abstract

This paper presents preprocessing algorithms for reducing misalignment of co-channel speech signals in order to improve the accuracy of adaptive speech separation. The use of adaptive decorrelation filtering as a front end processor in speech recognition systems improve performance when the input signals from the two channels are synchronized. However the performance of ADF is degraded when the co-channel signals are misaligned. We address this problem by proposing a pre-processor to the ADF with a view to align the co-channel signals. We have implemented two different algorithms for synchronizing the co-channel speech signals using cross-correlation. These algorithms improve the performance of speech recognition.