Audio Signal Extraction From Single-channel Based On Dictionary Learning

Audio signal is one of the most important sources of information for people. The study of the audio signal mainly includes extraction, separation, recognition and so on, and how to exactly extract the target source signal from the mixed signals has been a hot issue in the field of the audio signal processing. The audio signal extraction is referred to the procedure which recovers one or more than one interested source signals from the observed signals without knowing the number of source signals. According the number of observed signals, the problem can be divided into multi-channel, dual-channel or single-channel speech extraction problems, and the single-channel speech extraction problem is the challenge problem.We apply the sparse representation to the single-channel mixed speech extraction according the sparse characteristic which the natural signal exist in the transform domain. Firstly, we employ the characteristic that the speech signal will be extracted as the training signal for the analysis dictionary learning, and then the analytic dictionary can be obtained by using some dictionary learning algorithms. Then the dictionary subset which is orthogonal with the signal in the dictionary can be found by using the learning dictionary. Finally, the target signal could be extracted based on the nonorthogonality between the dictionary subset and the other interference signal.