Speech recognizer based on blind source separation using independent sparse strategies

Abstract

Many methods are used for blind sources separation, whereas it has non Gaussian and sample dependences properties, which fails to obtain the efficient entropy rate for the separation process. The first work of this dissertation describes the independent component analysis based on blind source separation using the Markovian and Invertible Filter Model. In order to obtain the efficient entropy rate, designed by utilizing the two types of source models are Markovian, and another one is the invertible filter model that gives the general Independent Component Analysis (ICA) uses mutual information rate for the analysis. Under the Markovian source model, the entropy rate equals the difference between two joint entropies. Under the invertible filter source model, the source is generated by an invertible filter driven independently and identically distributed random process. Therefore, the entropy rate of the source equals the entropy of the driving process under some constraints. Even though the combined strategy of the Markovian model with an invertible filter model produces a better estimation of entropy, there are some issues in terms of time complexity and explicit selection of signal is still lacking, which affects the separation of source signal from the mixture of signals. The second work of this dissertation describes an underdetermined blind speech signal separation. Thus above-stated issues can be solved by the Underdetermined Blind Speech Signal Separation (UBSS) problem when the number of observation is less than the sources for which the ICA is no longer applicable, which enhance the time complexity for separation of the signal. The tackling process is done with the aid of Improved Sparse Component Analysis (ISCA) is introduced to exploit the sparse nature of the TF domain, which adopts a two-step processing that contains mixing matrix estimation followed by separation of the source. This ISCA is based on fuzzy c-means with Particle Swarm Optimization (PSO) algorithm for mixed matrix Estimation. Our propos