Fixed-Point ICA based Speech Signal Separation and Enhancement with Generalized Gaussian Model
Rajkishore Prasad · Institutional Repositories DataBase (IRDB) · 2005
Speech slgnal separation and enhancement under blind setup is one of the challenglng areaS Of practical application･ Excellent solutions to tbese problems are always required for the spoken communication between man and machine in the real world･ The problem of speech separation arises in the presence of multiple speakers and that of enhancement pertains to reduce the effect of noise and other interfering slgnals･ In the real world applications these two problems are often occurring simultaneously and their solutions are urgently required in the development of full-fledged converBational interface･ The aims and scope of our work is also in the same context･ Recently, Blind Signal Separation (BSS) based on the lndependent ComponentAnalysis (ICA) has emerged as a potential engineering solution for speech separation problem･ Such algorithms work with the assumption of statistical independence of each sources and estimate original sources as the independent or least dependent components･ This thesisalso addresses development and application of ICA based algorithm for tbe blind separation of convoluted mixture of speecb, observed by a two element linear microphone array, under the over-determined situation･ The proposed ICA algorithm is based on the non-Gaussianization, by negentropy maximization, of the Time-Frequency Series of Speech (TFSS) signal.The functioning of ICA by non-Gaussianization is based on the heuristic idea of Central Limit Theorem (CLT) under which it happens that the mixed speech signals become more Gaussian tban tbe individual signal and thus by reversing tbe process of non-Gaussianization individual signals can be estimated with arbitrary scale and Permutation･ Under such a framework a costfunction is required to measure the lll L二ﾆｺ±事-However, for the proposed post-processing the knowledge of the level of residual noise present in the target speech is required and can be determined from the information about noise reduction done by the FDICA algorithm.However, this method is not blind as it requires orlglnal contribution of each source to each microphone.The experimental results show that the post processing by the MAPS estimator glVeS aPPreCiable improvements in the noise reduction.