Geo-Centric Language Models for Local Business Voice Search

Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the ACL, pages 389–396, Boulder, Colorado, June 2009. c⃝2009 Association for Computational Linguistics Geo-Centric Language Models for Local Business Voice Search Amanda Stent, Ilija Zeljkovi´c, Diamantino Caseiro and Jay Wilpon AT&T Labs – Research 180 Park Avenue Bldg. 103 Florham Park, NJ 07932, USA stent, ilija, caseiro, jgw@research.att.com Abstract Voice search is increasingly popular, espe- cially for local business directory assistance. However, speech recognition accuracy on business listing names is still low, leading to user frustration. In this paper, we present a new algorithm for geo-centric language model generation for local business voice search for mobile users. Our algorithm has several ad- vantages: it provides a language model for any user in any location; the geographic area covered by the language model is adapted to the local business density, giving high recog- nition accuracy; and the language models can be pre-compiled, giving fast recognition time. In an experiment using spoken business list- ing name queries from a business directory assistance service, we achieve a 16.8% abso- lute improvement in recognition accuracy and a 3-fold speedup in recognition time with geo- centric language models when compared with a nationwide language model. 1 Introduction Voice search is an increasingly popular application of speech recognition to telephony. In particular, in the last two years several companies have come out with systems for local business voice search (LBVS). In this type of application, the user pro- vides a desired location (city/state) and a business name, and the system returns one or more match- ing business listings. The most traditional LBVS applications are commercial 411 services, which are implemented as a speech-only two-exchange dialog such as the one in Figure 1. In this approach to LBVS, the speech recognizer (ASR) uses one gram- mar to recognize city/state, and then uses separate grammars for recognizing listings in each local area. This gives relatively high recognition accuracy. Advancements in ASR and search technology have made a more information retrieval-style LBVS feasible. In this approach, the ASR typically uses a large stochastic language model that permits the user to specify location and listing name or cate- gory together in a single utterance, and then sub- mits recognition results to a search engine (Natara- jan et al., 2002). This gives the user more flexibility to “say anything at any time”. However, in recent evaluations of one-exchange LBVS we have found that locations are recognized with much higher ac- curacy than listing names1. This may mean that the user has to repeat both location and listing several times (while in a traditional two-exchange interac- tion only one piece of information would have to be repeated). In effect, system developers have traded recognition accuracy for interaction flexibility, po- tentially increasing user frustration. Advances in mobile phone technology make it possible for us to combine the advantages of two- exchange and one-exchange LBVS. The newest smart phones come with global positioning system (GPS) receivers and/or with the ability to …

Content truncated to 500 words