the second layer binds the information. Analysis

Abstract. A sound source separation technique based on a bio-inspired neural network, capable of functioning in more than two-source mix-tures, is proposed. Separation results are compared with other proposed techniques in the literature using quantitative evaluation criteria. 1 The sound source separation problem In our life we are confronted to situation in which a mixture of sound sources is present in the environment and our goal is to extract one of the sources among others. While the auditory system may not always succeed in this goal, the range of situations in which recognition is possible in the presence of competing sources highlights the flexibility and robustness of human in speech perception. Here we propose a technique that roughly simulates the behavior of the auditory pathway. Our separation technique uses the Computational Auditory Scene Analysis [1]. 2 The proposed model An enhanced FIR Gammatone filterbank is used to mimic the behavior of the cochlea [2]. From the output of the cochlear channels two different anthropo-morphic maps are generated. The Cochleotopic / AMtopic and Cochleotopic/ Spectrotopic Maps, which try to mimic partially the behavior of the peripheral auditory pathway are generated. These maps are based on the reassigned FFT (Fast Fourier Transform) and envelope detection [3]. A two-layered network of spiking neurons is used to perform cochlear channel selection (Fig. 2) based on temporal correlation: neurons associated to those channels belonging to the same sound source synchronize. 2.1 Three-source sound source separation In our previous works, we applied our proposed model to two-source sound source separation [3] [2]. Most of other proposed models in the literature for ASA-based sound source separation deal only with two-source sound source separation (see [4] [5] [6] [7] [8]). One of the exceptions to this general tendency is the work b