Single-channel talker localization based on separation of the acoustic transfer function using hidden Markov model and its classification

抄録

This paper presents a talker localization method using only a single microphone, where phoneme hidden Markov models (HMMs) of clean speech are introduced to estimate the acoustic transfer function from the user's position. In our previous work, we proposed a Gaussian mixture model (GMM) separation for estimation of the user's position, where the observed speech is separated into the acoustic transfer function and the clean speech GMM. In this paper, we propose an improved method using phoneme HMMs for separation of the acoustic transfer function. This method expresses the speech signal as a network of phoneme HMMs, while our previous method expresses it as a GMM without considering the temporal phonetic changes of the speech signal. The support vector machine (SVM) for classifying the user's position is trained using the separated frame sequences of the acoustic transfer function. Then, for each test data set, the acoustic transfer function is separated, and the position is estimated by discriminating the acoustic transfer function. The effectiveness of this method has been confirmed by talker localization experiments performed in a room environment.