Co-Channel Speech Analysis

Abstract: Co-channel speech may be defined as a speech that is mixed with another speech. An example of a co-channel speech signal is generated when two or more people are speaking simultaneously (e.g. the cocktail-party. Processing the co-channel signal, i.e speaker identification, speech recognition, etc.. still have problems. Questions like how many speakers are talking together?, and if it is possible to decide for sure that a certain speaker is presented among the talking people, may need some answers. In the present work, using a predefined recorded speaker signal, target speaker, a proposed spectral correlation technique for both co-channel speech detection, and speaker identification is designed to help in resolving some of these problems.
This signal represents a great challenge for automatic speech applications. Performance of automatic speech recognition (ASR) and automatic speaker identification (ASI) systems , has been shown to degrade significantly in the presence of such signal.
If there is more than one speaker in the speech signal, co-channel speech, the speaker identification system will have some troubles to make a correct decision about the target speaker. So, most of the work done in speech recognition, and speaker identification assuming that there is one or two speaker speech signal(s) [1][2][3][4][5][6][10][[13]. A complete speech database like KED TIMIT contains 453 utterances spoken only by a US male speaker [7]. Thus, the goal of speaker identification system, when dealing with a single speaker signal is to say whether that speaker is the target speaker or not. When dealing with a co-channel speech signal, i.e multi-speaker signal, the identification system has been to identify the presence of the target speaker inside the signal [3][4]. If the co-channel speech signal has more than two speakers, three or more, the problem of speaker identification will be more difficult to
At first a literature preview about previous work concerning co-channel multi-speaker signal processing and speaker identification is illustrated. Secondly, the proposed system model is presented to describe the steps of problem solution and system operation. Then, the system is represented mathematically and its algorithms are designed and tested. At last, speaker signal files which are used in the experimental work, collected from Corbus database and some Arabic recorded files are used to test the algorithms and system. Matlab programming environment is utilized to design and run the system algorithms. The output results are plotted and tabulated using Matlab program. Conclusions are summarized and stated at the end of the paper. Future work concerning the point of research is also