Estimation of distance of a target speech source by involving monaural features and statistical properties

Abstract

The paper discusses a novel system for the estimation of distance of a target speaker by involving statistical properties in a reverberant condition. The system involves the extraction of statistical features from both cepstral and envelope coefficients of a speaker at different distances. Further, different spectral or monaural features are analysed at distinct distances for different room environments. The distance-dependent statistical properties are considered for the feature extraction process. A set of statistical parameters are used to learn GMM-EM pattern recognizer for effective classification. The results observed that the system performance is very much dependent on the reverberation time and also robustness of the monaural features. The results of the proposed system show the significant improvement in signal-to-noise ratio of 0 dB (babble noise) under reverberation time 0.48 s over other existing methods.