A New Model for Robot Audition using Independent Component Analysis and Time-Frequency Representation

We address the problem of robot audition and propose an approach to identify the speakers in this paper. The independent component analysis and time-frequency representation are introduced into the auditory research, and an auditory system is built on this basis. We employ the improved deconvolutive independent component analysis to separate the time delayed and convolved speech signals which are received by robots. The results of independent component analysis are inherently out-of-order, so robots cannot correlate the separated signals with sources. To solve this problem, we exploit the combination of optimized time-frequency representation and distance measures to reorder them, and accomplish the identification of speech signals. In identifying experiments, the best-obtained false acceptance rate is 0.15%, and the false rejection rate is 0.1%. The result reveals that we have implemented the auditory function of automaton with a good performance by the presented system.