Towards Using Prosody in Speech Recognition/Understanding Systems: Differences Between Read and Spontaneous Speech

TOWARDS USING PROSODY IN SPEECH RECOGNITION/UNDERSTANDING SYSTEMS: DIFFERENCES BETWEEN READ AND SPONTANEOUS SPEECH Kim E.A. Silverman, Eleonora Blaauw l, Judith Spitz, John F Pitrelli Speech Technology Group Artificial Intelligence Laboratory NYNEX Science and Technology White Plains, NY, 10604, U.S.A. 1. Ph.D. student in Institute for Language and Speech, University of Utrecht 1. ABSTRACT A persistent problem for keyword-driven speech recognition sys- tems is that users often embed the to-be-recognized words or phrases in longer utterances. The recognizer needs to locate the relevant sections of the speech signal and ignore extraneous words. Prosody might provide an extra source of information to help locate target words embedded in other speech. In this paper we examine some prosodic characteristics of 160 such utterances and compare matched read and spontaneous versions. Half of the utterances are from a corpus of spontaneous answers to requests for the name of a city, recorded from calls to Directory Assistance Operators. The other half are the same word strings read by volun- teers attempting to model the real dialogue. Results show a consis- tent pattern across both sets of data: embedded city names almost always bear nuclear pitch accents and are in their own intonational phrases. However the distributions of tonal make-up of these pro- sodic features differ markedly in read versus spontaneous speech, implying that if algorithms that exploit these prosodic regularities are trained on read speech, then the probabilities are likely to be incorrect models of real user speech. 2. INTRODUCTION This work addresses two related questions. One is whether spontaneous goal-directed utterances collected from real users in a particular application domain exhibit reliable pro- sodic patterns that could be exploited by recognition algo- rithms. We focus on to-be-recognized words that are spoken within longer utterances, in order to investigate whether these embedded words have particular prosodic characteristics that could help a recognizer to locate them. One of the original motivations for this study was our observation from informal listening to our corpus that such embedded words bear nuclear pitch accents. If this is a con- sistent pattern, it would mean that they are (1) louder, longer and more clearly articulated than they would be without nuclear accents, and (2) they would bear character- istic fundamental frequency movements. Corpora of spontaneous goal-directed speech from real users are not readily obtainable, and so it is common prac- tice to record speech read out by volunteers in order to develop, train and test recognition algorithms. To the extent that the prosody of read speech differs from that of sponta- neous goal-directed speech, such "laboratory" corpora may obscure or misrepresent any reliable prosodic properties found in spontaneous "real user" speech. Consequently the second question investigated in this work is whether such patterns can also be found in recordings of speech reead out by volunteers. We are interested in prosodic differences between read and spontaneous speech because of their rele- vance to speech recognition, and for increasing the natural- ness of synthetic speech. It is worth pointing out a methodological …

Content truncated to 500 words