Exploiting Temporal Correlations in Joint Multichannel Speech Separation and Noise Suppression Using Hidden Markov Models

Abstract:
Recently, we introduced a unified approach for blind source separation and speech enhancement using microphone arrays [1]. Based on a sparseness model for the observations we derived an expectationmaximization (EM) algorithm for blind system identification and dominant source activity detection. In this contribution we improve this approach in the following two aspects: Firstly, we extend the observation model to a hidden Markov model (HMM) to explicitly exploit temporal correlations. Secondly, we include power based speech presence probability estimation. These modifications result in significantly improved spectrograms of the speaker presence probabilities after source separation and improved performance with respect to interference, noise and artefact reduction of the overall system, as measured on the dataset of the Signal Separation Evaluation Campaign 2010. Index Terms — Microphone array, blind source separation, speech enhancement, hidden Markov model