Speech and Vision Lab

In automatic speech recognition systems, the information in the speech signal is traditionally retrieved in the form of feature vectors representing sub-word units and thereby converting the features into human readable text form. However, these systems perform poorly due to degradations of speech under varying environmental conditions. To improve the performance, the main issues to be considered are: (a) Determination of speech regions in the speech data collected in degraded environments, and (b) Recognition of speech sounds from the degraded speech in the detected speech regions. Although there exist wide variety of techniques which address these issues, most of them are applicable for clean speech synthetically degraded by stationary noise conditions, due to the need for large amount of training data for statistical modeling. The present work focuses on methods of processing the signals so as to determine the desired speech regions in degraded conditions. For this, signal processing methods are being explored to extract speech-specific characteristics independent of the characteristics of degradations.