The group delay function has been used conventionally in temporal spectral analysis and feature extraction for speech recognition. In this work we present a detailed analysis of a novel approach to spatial spectral analysis of speech using the MUSIC-Group delay spectrum. In our previous work we have proposed the use of the MUSIC-Group delay spectrum [ICASSP 2010], for direction of arrival estimation (DOA) and distant speech recognition. We discuss the advantages of the proposed method in terms of resolving closely spaced speech sources with minimal number of sensors. This method is also analyzed from a minimum phase perspective as is done in temporal processing of speech. Additional analysis is performed using the Pisarenko-Group delay spectrum in terms of real time performance. DOAs estimated from the proposed approach are used to train filter and sum beamformers. Distant speech recognition experiments in clean and reverberant conditions using the beamformed speech signal indicate reasonable improvements over correlation and sub space based methods.
展开▼
