February Fourier Talks 2007

Shubha Kadambe


Underdetermined convolutive mixture separation and its application to robust automatic speech recognition


Mixed signals that are received at the sensors like the microphone and antenna cause degradation in the performance of automatic speech recognition and cellular communication. Hence, it is essential to separate the mixed signals. Generally, the mixing environment (system) is unknown and also the source signals that are received by the sensors. Therefore, blind techniques are needed to separate the mixed source signals. Mainly one has to deal with two types of mixtures - instantaneous and convolutive. Convolutive mixture separation is a harder problem to solve but more practical problem to solve. Several blind techniques have been developed for both instantaneous and convolutive mixture separation. Most of these techniques assume that the number of sensors are equal to the number of sources. This assumption is impractical. So, we have developed a probabilistic technique that is applicable for mixture separation when the number sensors is less than the number of sources. This technique jointly minimizes norm 1 and 2 to estimate the mixing system parameters and the source signals simultaneously in an iterative fashion. In this talk, a detailed description of this technique will be provided. We have applied this technique for the separation of various types of signals - speech, communication and radar. In this talk, the application for the separation of speech signals will be provided. We will also demonstrate the performance improvement in the automatic speech recognition by 15-30% after applying our technique to enhance the speech signals before extracting the features that are used by the automatic speech recognizer.