Integrating Prosodics into a Language Model for Spoken Language Understanding of Thai

Integrating Prosodics into a Language Model for Spoken Language Understanding of Thai* Siripong Potisuk Department of Electrical and Computer Engineering The Citadel, the Military College of South Carolina 171 Moultrie Street Charleston, South Carolina 29409 USA siripong.potisuk@citadel.edu Abstract. This paper describes a preliminary work on prosody modeling aspect of a spoken language understanding system for Thai. Specifically, the model is designed to integrate prosodics into a language model based on constraint dependency grammar. There are two steps involved, namely the prosodic annotation process and the prosodic disambiguation process. The annotation process uses prosodic information (stress, pause and syllable duration) obtained from the speech signal during low-level acoustic processing to encode each word in the parse graph with a prosodic feature called strength dynamic. The goal of the annotation process is to capture the correspondence between the phonological and phonetic attributes of the prosodic structure of utterances and their intended meanings. The prosodic disambiguation process involves the propagation of prosodic constraints designed to eliminate syntactic ambiguities of the input utterance. It is considered a pruning mechanism for the parser by using prosodic information. The aforementioned process has been tested on a set of ambiguous sentences, which represents various structural ambiguities involving five types of compounds in Thai. Keywords: Spoken language understanding, Prosody, Thai. 1. Introduction In the realm of artificial intelligence, there is no denying that the ultimate goal for man-machine communication is to use natural human languages. Humans aspire to build a machine capable of understanding and responding to commands issued in the form of spoken language. Research towards this goal encompasses various research disciplines, and the fruits of such labor include machine translation, speech synthesis and recognition, and spoken language understanding. Spoken language understanding has been a challenge for scientists and engineers for many decades. Early efforts were primarily aimed at simply recognizing speech, and extensive research has been conducted to improve performance of speech recognition systems. Much of the improvements can be attributed to combining various research disciplines including linguistics, psychology, computer science, and signal processing, as well as improvements in modeling techniques, such as statistical and symbolic pattern recognition. Speech models now incorporate linguistic modeling to account for coarticulation and grammatical constraints to reduce the search * The author would like to thank the Citadel Foundation for its financial support in the form of a presentation grant. Copyright 2008 by Siripong Potisuk 311 22nd Pacific Asia Conference on Language, Information and Computation, pages 311–320 space for the correct utterance. Also, steady advances in the computational efficiency of computers and hardware have enabled researchers to implement computationally intensive models that were deemed nearly impossible in the past. With the remarkable advances in speech recognition, the focus has now shifted toward the development of spoken language understanding systems. The goal is to build a machine capable of understanding naturally spoken language by combining together speech recognition and natural language processing technologies. Such systems must be capable of recognizing a large number of words (on the order of tens of …

Content truncated to 500 words