TOWARDS ROBUST SPEECH RECOGNITION IN THE TELEPHONY NETWORK ENVIRONMENT - CELLULAR AND LANDLINE CONDITIONS

Abstract

We describe several speaker-independent speech recognition studies conducted with both landline and cellular network telephone data. The cellular environment included the three dominant standards found in the United States: CDMA, TDMA and GSM. Our goal was to design a system that operated over all these four channels, handling their innate variations, such as those of background and line characteristics. Our baseline system was trained on 200 hours of landline telephone speech from about 25,000 speakers. We experimented with MAP adaptation utilizing some training sentences from our cellular and landline databases. We applied an LDA procedure to improve our performance. We compared the performances of these systems on several independent test databases and demonstrated the effectiveness of a hybrid system built with data taken from all four networks.