A speech understanding system based on a topic-oriented language model
Masahiro Hori, Katsuhiko Tsujino, Riichiro Mizoguchi, Osamu Kakusho · The Journal of the Acoustical Society of America · 1988
Higher-level knowledge plays an important role in the speech understanding problem. A topic-oriented language model is used in SPURT-I (a SPeech Understanding System with Rule-based and Topic-directed architecture). SPURT-I consists of three subsystems: a rule-based acoustic analyzer SPREX (a SPeech Recognition EXpert), a word hypothesizer, and a topic-directed parser ASP (an ASsociated-based Parser). SPURT-I accepts Japanese speech divided into “bunsetus” (Japanese syntactic units), and attempts to output a corresponding word sequence. The basic assumption of this approach is that acoustically close phoneme sequences rarely correspond to semantically close words. Based on this assumption, SPREX recognizes a group of consonants such as /b,d,g,z/, /p,t,k/, and so on. As a result of this simplification, more candidate words are generated by the word hypothesizer than in the case of strict recognition. Then ASP reduces the effective number of candidates by associating the topic of utterance and the candidate words. Currently, SPURT-I has a 1000-word vocabulary for simple scenic descriptions. Experiments with actual utterances showed that nine sentences out of ten were recognized successfully and the “bunsetu” recognition rate was 97%. [Work partly supported by a Grant-in-Aid for Scientific Research, Ministry of Education.