Using a low-bit rate speech enhancement variable post-filter as a speech recognition system pre-filter to improve robustness to GSM speech

Master Thesis


Permanent link to this Item
Journal Title
Link to Journal
Journal ISSN
Volume Title

University of Cape Town

Performance of speech recognition systems degrades when they are used to recognize speech that has been transmitted through GS1 (Global System for Mobile Communications) voice communication channels (GSM speech). This degradation is mainly due to GSM speech coding and GSM channel noise on speech signals transmitted through the network. This poor recognition of GSM channel speech limits the use of speech recognition applications over GSM networks. If speech recognition technology is to be used unlimitedly over GSM networks recognition accuracy of GSM channel speech has to be improved. Different channel normalization techniques have been developed in an attempt to improve recognition accuracy of voice channel modified speech in general (not specifically for GSM channel speech). These techniques can be classified into three broad categories, namely, model modification, signal pre-processing and feature processing techniques. In this work, as a contribution toward improving the robustness of speech recognition systems to GSM speech, the use of a low-bit speech enhancement post-filter as a speech recognition system pre-filter is proposed. This filter is to be used in recognition systems in combination with channel normalization techniques.

Includes bibliographical references.