This paper describes our first attempt to model accents in Japanese. We chose connected-digit recognition as our task since its simple accent structure will easily enable us to model its accents. We propose to use two distinct models for each digit according to two possible accent positions. Analysis of the trained models seems to show that the models represent alterations in the acoustics introduced by accents, even though current models only utilize intensity to distinguish accent. Connected digit recognition tests show that word errors can be decreased by up to 42% with the proposed models compared to conventional accent-independent models. Error analysis seems to suggest that not all digits require accent-dependent models.
Bibliographic reference. Kondo, Kazuhiro (1995): "Connected Japanese digit recognition with pitch accent-dependent models", In EUROSPEECH-1995, 23-26.