An Online Expectation-Maximization Algorithm for Tracking Acoustic Sources in Multi-Microphone Devices During Music Playback

摘要

In this paper, we propose an expectation-maximization algorithm to perform online tracking of moving sources around multi-microphone devices. We are particularly targeting the application scenario of distant-talking control of a music playback device. The goal is to perform spatial tracking of the moving sources and to estimate the probability that each of these sources is active. In particular, we use the expectation-maximization algorithm to capture the statistical behavior of the feature space representing the ensemble of sources as a Gaussian mixture model, assigning each Gaussian component to an individual acoustic source. The features used exploit a wide range of information on the sources behavior making the system robust to noise, reverberation, and music playback. We then differentiate between desired and interfering sources. The spatial information and activity level is then determined for each desired source. Experimental evaluation of a real acoustic source tracking problem with and without music playback shows promising results for the proposed approach.