One source signal extraction based on metrics transform

A new approach for one source signal extraction based on metrics transform is proposed and investigated in this paper. First, the mixing matrix is estimated by employing the K-means algorithm on single-source-points with higher energies. Then, the time-frequency points that incorporate one source, which can be used to approximately denote the source signals that include main information of them, are found out by employing a novel metrics transform separation algorithm. Next, the Mel Frequency Cepstral Coefficients of these signals as well as the referenced signal are extracted respectively. And here we will get the index of the target signal which has the maximum similarity to the referenced signal. After that, we apply the metrics transform step by step to find the time-frequency points that incorporate two sources, three sources and so on. A key point is that only those points which contain the contributions of the target signal will be processed. Finally the target signal is obtained through the inverse short-time Fourier transform. Compared with existing methods, our approach can be used even for the case in which the number of mixtures is smaller than that of sources and does not need any extra process after the separating. Experimental results indicate the validity of the method.