As speech has been expected to serve as one of the most natural media in man-machine communication, speech recognition technology has been continuously improved year after year. At the present time, speech recognition technology has migrated from main-frame computers to work-stations and even personal computers, and applications are already running on them. There are several factors that have contributed to this rapid progress: (1) the development of large speech corpora, (2) the standardization of performance evaluation, and (3) the advances in computer technology.
Currently, a statistical method based on the Hidden Markov Models (HMMs) is well-known and widely used in most current research systems. HMM is powerful in that, with the availability of training data, the parameters of the model can be adjusted automatically to give optimal performance. The current status of this approach related to Automatic Speech Recognition (ASR) for Japanese language is briefly described.
