Blind Source Separation Based On Independent Component Analysis And Pitch Detection

The subject of this article is about the Blind Source Separation (BSS). So far, Independent Component Analysis (ICA) is the most important method by which the researchers can deal with the blind separation of signals. The purpose of ICA is to reach a seperate matrix which allows the separated signals to be statistically independent each other. 1 lowever, the sequence of separated signals will not bring effects on the statistical independence of those signals. For this reason the algorithm of ICA has many equivalent solutions, that is, the sequence of separated signals will give rise to all kinds of corresponding possibilities. But it is the uncertainty of this sequence that is just the main disadvantage of ICA, because the sequence of signals is usually critical in the real conditions.When we try to solve the blind separation of speech signals in some specific occasions, it is to be noted that the speech signals themslves contain lots of information. If we can acquire the sequence information from the original information of signals, the above problem about ICA can be solved.The pitch is one of the primary properties of the speech signals. The pitch is charactered as quasi-stationary, and can be referred as a parameter which reflects the frequency and energy qualities of the speech signals. There are two questions when using the pitch information of the speech signals to deal with the sequence problem: one is how to extract the pitch from the mixing speech signals; the other is how to employ sufficiently the pitch information to solve the sequence problem.To solve the former, this article, based on the sine model of the speech signals, develops a new algorithm to extract the pitch from the speech signals by analysing the average energy value of pitch frequency harmonic. The algorithm calculates the average energy value and describes it by the figures, then finds all of the extreme points, and inturn extracts the pitch frequency from the mixing signals by the filters of frequency-doubling and frequency-dividing.To solve the latter, this article compares the frequency domains of the original signals with those of the separated signals. By connecting the harmonic forms of the pitch frequency with the above comparison together, this article establishes a sequencable algorithm which adopts simultaneously the separation matrix and the pitch information.After overcoming the above difficulties, we design a sequential blind separation system based on the pitch information, and analyse and verify this system experimently.Finally, we use the DSP processor TMS320C5410 produced by the TI company to present the software and hardware design for the mixing speech acquisition and the blind separation system, and provide some source codes.