Improving the Performance of Automatic Speech Recognition Using Blind Source Separation

Abstract

In real world applications, Speech recognition system have grown due its significance in various online and offline applications such as security, robotic application, speech translator etc. These systems are widely used now-a-days where acquisition of signal is performed using various instruments which causes noise, source mixing and other impurities which affects the performance of speech recognition system. In this work, issue of source mixing in original speech signal is addressed which causes performance degradation. In order to overcome this we propose a new approach which utilizes non-negative matrix factorization modelling. This method utilizes scattering transform by applying wavelet filter bank and pyramid scattering to estimate the source and minimization of unwanted signals. After estimation the signals or sources, source separation algorithm is implemented using optimization process based on the training and testing method. Proposed approach is compared with other existing method by computing performance measurement matrices which shows the better performance