Speech recognition for a travel reservation system

In this paper, we present our work on speech recognition for a spoken dialog system for automatic travel reservations. The system can receive speech input from an analog line or from IP telephony through a web site. The system uses speech recognition, natural language understanding, air travel database access, dialog management, natural language generation and speech synthesis technologies to perform the task of an automated travel agent. Language modeling for a mixed-initiative spoken dialog system is a challenging problem. We explore class-based n-gram language models (LMs) for this domain. Details for designing LM classes and assigning non-uniform probabilities within the classes are provided. We point out disadvantages of class-based LMs and introduce compound word vocabularies to overcome the problems observed. Dialog state dependent LMs are explored which enables incorporating semantic context into the language models. We also analyze the use of embedded context free grammar objects within LMs and point out advantages and disadvantages of using them. The resulting LMs are implemented and evaluated using IBM telephony speech recognition engine. Our efforts are shown to decrease the word error rate from 24% to 17% on an evaluation testset.