Single-channel multi-talker-localization based on maximum likelihood

This paper presents a sound source (talker) localization method using only a single microphone based upon maximum likelihood. In our previous work, we proposed GMM (Gaussian mixture model) separation for estimation of the sound source direction, where the observed (reverberant) speech is separated into the acoustic transfer function and the clean speech GMM, and showed its effectiveness for the single-talker localization task. In this paper, we discuss a multi-talker localization method using GMM separation and model composition. Model composition is used to represent speech signals observed in a reverberant environment corresponding to every conceivable combination of positions of the sound sources, where composite models are obtained through composition of talker´s speech model and acoustic transfer functions estimated using GMM separation. For each test data set, we find a maximum-likelihood model from among the composite models corresponding to each combination of talkers´ positions. The effectiveness of this method has been confirmed by two-talker localization experiments performed in a room environment.