A Method for Open-Vocabulary Speech-Driven Text Retrieval

A Method for Open-Vocabulary Speech-Driven Text Retrieval Atsushi Fujii � University of Library and Information Science 1-2 Kasuga, Tsukuba 305-8550, Japan fujii@ulis.ac.jp Katunobu Itou National Institute of Advanced Industrial Science and Technology 1-1-1 Chuuou Daini Umezono Tsukuba, 305-8568, Japan itou@ni.aist.go.jp Tetsuya Ishikawa University of Library and Information Science 1-2 Kasuga, Tsukuba 305-8550, Japan ishikawa@ulis.ac.jp Abstract While recent retrieval techniques do not limit the number of index terms, out-of- vocabulary (OOV) words are crucial in speech recognition. Aiming at retrieving information with spoken queries, we fill the gap between speech recognition and text retrieval in terms of the vocabulary size. Given a spoken query, we gener- ate a transcription and detect OOV words through speech recognition. We then cor- respond detected OOV words to terms in- dexed in a target collection to complete the transcription, and search the collection for documents relevant to the completed tran- scription. We show the effectiveness of our method by way of experiments. 1 Introduction Automatic speech recognition, which decodes hu- man voice to generate transcriptions, has of late become a practical technology. It is feasible that speech recognition is used in real-world human lan- guage applications, such as information retrieval. Initiated partially by TREC-6, various methods have been proposed for “spoken document retrieval (SDR),” in which written queries are used to search speech archives for relevant information (Garo- folo et al., 1997). State-of-the-art SDR methods, where speech recognition error rate is 20-30%, are � The first and second authors are also members of CREST, Japan Science and Technology Corporation. comparable with text retrieval methods in perfor- mance (Jourlin et al., 2000), and thus are already practical. Possible rationales include that recogni- tion errors are overshadowed by a large number of words correctly transcribed in target documents. However, “speech-driven retrieval,” where spo- ken queries are used to retrieve (textual) informa- tion, has not fully been explored, although it is re- lated to numerous keyboard-less applications, such as telephone-based retrieval, car navigation systems, and user-friendly interfaces. Unlike spoken document retrieval, speech-driven retrieval is still a challenging task, because recogni- tion errors in short queries considerably decrease re- trieval accuracy. A number of references addressing this issue can be found in past research literature. Barnett et al. (1997) and Crestani (2000) indepen- dently performed comparative experiments related to speech-driven retrieval, where the DRAGON speech recognition system was used as an input in- terface for the INQUERY text retrieval system. They used as test queries 35 topics in the TREC col- lection, dictated by a single male speaker. How- ever, these cases focused on improving text retrieval methods and did not address problems in improv- ing speech recognition. As a result, errors in recog- nizing spoken queries (error rate was approximately 30%) considerably decreased the retrieval accuracy. Although we showed that the use of target docu- ment collections in producing language models for speech recognition significantly improved the per- formance of speech-driven retrieval (Fujii et al., 2002; Itou et al., 2001), a number of issues still re- main open …

Content truncated to 500 words