Prosody and Speech Recognition

chart parsing is reduced into a set of matrix operations dealing with sparse matrices. Appendices list sample gram­mars and lexicons, which brings substance to the claims.

"Speech" in this book refers to English only, which is never made explicit. This seems to be the normal case in American literature, however. Of course, most of the contri­bution is relevant to other languages as well.

The book also provides an interesting contribution in the area of finite-state properties of language, because the phrase structure grammars used are essentially finite-state. Other finite-state accounts (such as the two-level model by Koskenniemi [1984] and cascaded transducers by Kaplan and Kay [Kay 1983] seem to have been less successful in combining structural information with segmental pro­cesses. Both other models are purely segmental, although syllables are sometimes referred to as contexts.

An interesting problem concerning rule interaction in the proposed formalism is dealt with on page 113. There would be an obvious need for subtraction (for defining negative contexts) and intersection (combining effects). Subtrac­tion, however, turns out to exclude too much, whereas intersection is too permissive.

The book is well written and the argumentation proceeds logically. Both strong and weak points of the theories proposed are clearly presented. It gives a fair overall pic­ture of the field of speech recognition, and much of the book could be suitable as a textbook. Nevertheless, some pas­sages address mostly readers with a considerable back­ground. The main topic covers, of course, a specific slice of the whole field, namely the treatment of allophonic varia­tion. One minor inconvenience is the use of a reference format that cites a number only, not the author and the year. This results in a small savings in space but a larger burden for the reader. The book appears to be Church's (previously unpublished) doctoral dissertation from MIT, though this is not clearly indicated in the volume. Although not particularly new, it still is very valuable.

Kay, M. 1983 "When Meta-Rules Are not Meta-Rules," In Sparck Jones, K. and Wilks, Y. (eds.), Automatic Natural Language Parsing. Ellis Horwood, Chichester, U.K.

Koskenniemi, K. 1984 "A General Computational Model for Word-Form Recognition and Production," In COLING '84: Proceedings of the 10th International Conference on Computational Linguistics, Stanford, 178— 181.

Kimmo Koskenniemi's Ph.D. thesis introduced a finite-state two-level model to account for inflection, derivation, and compounding in an efficient and language-independent way. He works on finite-state morphology and syntax at the Research Unit for Computational Linguistics at the University of Helsinki. Kosken-niemi's address is: Department of General Linguistics, University of Helsinki, Hallituskatu 11, SF 00100 Helsinki, Finland. E-mail: koskenniemi@opmvax.csc.fi

Prosody and Speech Recognition AlexWaibel

(Carnegie Mellon University) Reviewed by Joan Bachenko AT&T Bell Laboratories

Most research in natural language processing (NLP) con­centrates on the syntax and semantics of written language, a situation that exists in part because most NLP applica­tions are concerned with systems that rely on written language analysis, e.g. information retrieval and text-generation systems. Recently, however, we have begun to see a growing interest in spoken language and the applica­tion of natural language processing to text-to-speech synthe­sis and speech recognition. Waibel's volume, which de­scribes new results in automated speech recognition, makes an important contribution to this research direction.

At present, speech recognition technology gives us two choices: speaker-independent systems that handle small vocabularies (one to five phonetically distinct words) and require no training, or speaker-dependent systems that recognize somewhat larger vocabularies (up to 1,000 words online) and require training sessions for each user. Al­though experimental systems can recognize limited contin­uous speech, freely generated phrases and sentences cannot be processed, nor can words that are "unknown" to a recognizer. Waibel believes that this condition can change if recognition systems, which currently focus on identifying acoustic phonetic segments, are expanded to include pro­sodie information, i.e. information about nonsegmental features such as duration and pitch. His central claim is that a system equipped with prosody rules can achieve very large vocabulary recognition, up to 20,000 words, in both continuous speech and isolated word tasks. To make his point, Waibel examines four prosodie features—duration, intensity, pitch, and stress. For each feature, he discusses in detail a series of experiments that demonstrate the tech­niques that he used (and, in some cases, invented) for extracting the prosodie features, and rules that use prosodie feature patterns to narrow the search space for word hypoth-esization to a small subset of the total vocabulary. Waibel's results make a convincing case that prosody can play a valuable role in machine perception of speech; however, they fall short of establishing his strongest claims, as he lacks a complete implementation of the system.

Most of the book is organized around each of the pro­sodie features that Waibel investigates. His explication is