Language Identification via Large Vocabulary Speaker Independent Continuous Speech Recognition

Language Identification via Large Vocabulary Speaker Independent Continuous Speech Recognition Steve Lowe, Anne Demedts, Larry Gillick, Mark Mandel, Barbara Peskin Dragon Systems, Inc. 320 Nevada Street Newton, Massachusetts 02160 ABSTRACT The goal of this study is to evaluate the potential for using large vocabulary continuous speech recognition as an engine for automatically classifying utterances according to the lan- guage being spoken. The problem of language identification is often thought of as being separate from the problem of speech recognition. But in this paper, as in Dragon's earlier work on topic and speaker identification, we explore a uni- fying approach to all three message classification problems based on the underlying stochastic process which gives rise to speech. We discuss the theoretical framework upon which our message classification systems are built and report on a series of experiments in which this theory is tested, using large vocabulary continuous speech recognition to distinguish English from Spanish. 1. INTRODUCTION In this paper we describe preliminary work being con- ducted at Dragon Systems exploring the use of large vo- cabulary continuous speech recognition as an engine for automatically classifying spoken utterances by language. Several approaches to the problem of language identifi- cation have already appeared in the literature, but they generally address the problem as quite separate from the problem of speech recognition. For example, LIMSI [1] has reported results in French-English discrimination via phone recognition and a number of sites, such as OGI [2], have performed language classification by using broad phonetic labels and analyzing sets of phonological "fea- tures". Our approach to the problem of language identification grows naturally out of our model for the underlying stochastic process giving rise to speech. In earlier pa- pers ([3], [4]) we have described our unified approach to the problems of topic and speaker identification via large vocabulary continuous speech recognition and demon- strated the success of this strategy even in classifying speech data in domains where the recognition task is far too difficult to obtain accurate transcriptions. We be- lieve that the contextual information - both acoustic and language model - available in full-scale large vocabulary continuous speech recognition is invaluable in extracting reliable data from difficult speech channels. We now ex- amine how this same framework supports work on the problem of language identification. In the next section we describe the theoretical founda- tions upon which our message classification systems are based and discuss some simplifying approximations in- troduced in their implementation. We then describe our initial testing of English-Spanish discrimination, primar- ily work with microphone data using our Wall Street Journal speech recognition system, but also work we are now beginning in language identification on telephone speech. Finally, we discuss some lessons learned from these early explorations and suggest plans for future work. 2. THEORETICAL FRAMEWORK We briefly review the theoretical background described in our earlier papers [3] and [4]. Our approach to the message classification problem - for topic, speaker, or language identification - is based on modelling speech as a stochastic …

Content truncated to 500 words