Dual-Type Automatic Speech Recogniser Designs for Spoken Dialogue Systems

Proceedings of the Australasian Language Technology Workshop 2005, pages 78–86, Sydney, Australia, December 2005. Dual-Type Automatic Speech Recogniser Designs for Spoken Dialogue Systems Jason Littlefield and Michael Broughton Command and Control Division, Defence Science and Technology Organisation PO Box 1500 Edinburgh, Australia, 5111 Jason.Littlefield@dsto.defence.gov.au Michael.Broughton@dsto.defence.gov.au Abstract A Dual-Type automatic speech recogniser (ASR) is a multi-pass ASR system that incorporates both a speaker-independent (SI) and a speaker-dependent (SD) ASR. The purpose of this approach is to improve the robustness of spoken dialogue systems for a broader range of applications. This paper identifies feasible Dual-Type multi-pass ASR system designs that are intended to overcome limitations arising from the use of a single type of ASR. Implementation issues are also discussed. 1.Introduction Current implementations of Spoken Dialogue Systems (SDS) are developed around a single ASR. This design limits the overall system’s recognition accuracy to the performance of the installed ASR, while the useability is determined by the individual ASR type. ASRs can be categorised into two types, either SD or SI, with each having their own strengths and weaknesses. SI ASRs have the advantage of not requiring a prior enrolment or customised training session for their end users, thereby allowing any user of a given regional dialect to effectively use the system. These systems rely on an underlying grammar that typically needs to be relatively small, or at least, only have a small portion of the grammar active at any point in time. Due to requiring limited grammar size for optimum recognition accuracy, these systems are often used in a system led interaction, whereby the machine asks questions of the user that elicit simple responses. These responses may be single words, or small continuous strings. SD ASRs, on the other hand, require training for each individual user. This training is relatively short, generally less than ten minutes, and involves the speaker reading aloud a prepared text, which is analysed by the ASR for generation of the speaker’s acoustic model. A speaker profile is created by combining this acoustic model with a vocabulary and a regional language model. SD systems are adaptive and have the advantage of being able to recognise free speech or constrained grammar speech, although extracting semantic content from the free speech is more difficult than with a formalised grammar. In addition to the fundamental differences between the two ASR types, there are also individual differences within each ASR type. The various commercial and research implementations are developed from different algorithms and techniques, typically providing varying output for the same paragraph of spoken text. ASRs are also further classified by their speech continuity as well as grammar and vocabulary size. Speech continuity describes whether words are spoken in isolation, as connected speech or as continuous speech (Zue et al., 1997). Connected speech ASRs require pauses between multiple word phrases, whereas continuous speech ASRs do not. The grammar and vocabulary size refers to the number of phrases and words that can be spoken and recognized. A grammar can be characterised by the …

Content truncated to 500 words