BibTeX:

@article{IJIRSTV2I12120,
     title={A Wavelet based ICA Algorithm for Separation of linearly mixed Speech Signals},
     author={Mohit Kumar and Asst. Prof Saranjeet },
     journal={International Journal for Innovative Research in Science & Technology},
     volume={2},
     number={12},
     pages={292--296},
     year={},
     url={http://www.ijirst.org/articles/IJIRSTV2I12120.pdf},
     publisher={IJIRST (International Journal for Innovative Research in Science & Technology)},
}

In this paper, an algorithm is purposed which can be used for separation of linearly mixed speech signals. Initially, this task was done by a popular algorithm called Independent Component Analysis (ICA). ICA is an effective algorithm to decompose mixed in to independent components. But it is observed that, when number of independent speaker increases, ICA is not able to extract the independent components completely. Hence the performance of ICA is satisfactory only when two individual speakers are present. The problem of extracting three speech signals more accurately is solved by using a Discrete Wavelet Transform based ICA. As a result of wavelet decomposition, wide band signals are converted into narrow band which decreases the probability of finding two independent signals in same frequency band. Recorded signals by the microphones are fed to an analysis filter bank which decomposes the signals in to approximation and detail frequency bands. These two groups of signals are further processed by ICA separately and then combined by synthesis filter bank to get the independent source signals. Further statistical parameters of the DWT based ICA are compared with conventional ICA on the basis of their performance.

ICA, Discrete Wavelet Transform, Signal Separation, Non-Gaussian mixture, Mutual Information