ACOUSTIC MODELING IMPROVEMENTS IN A SEGMENT-BASED SPEECH RECOGNIZER

Abstract

In this paper we report on some recent improvements on the acoustic modeling in a segment-based speech recognition system. Context-dependent segment models and improved pronunciation modeling are shown to reduce word error rates in a telephone -based, conversational system by over 18%, while the technique of Gaussian selection reduces overall computation by more than a factor of two. 1. INTRODUCTION Since its deployment in 1997 [1], the number of calls made to the Jupiter telephone-based conversational weather information system has been steadily increasing. Currently, an average of about 200 calls, yielding 1000 new utterances, are made each day. This continuous influx of real data from a wide variety of users and channel conditions is an invaluable resource for our research in robust speech recognition and understanding. In this paper we report on recent refinements in acoustic modeling which have improved performance, and reduced computational requirements. The SUMMIT segmen...