Relationship Among Speaking Style, Inter-Phoneme's Distance and speech Recognition Performance

摘要

There is a limit of recognition performance for dialogue speech using acoustic models built only with read speech, because various acoustic and linguistic phenomena, whihc reflect the characteristics of spontaneous speech, are observed i nthe dialogue speech. In this paper, we investigated the differences of acoustic properties which cause the limit among isolated words, read speech and spontaneous speech. Firstly, the dialogue speech was compared with the read speech through acoustic analyses. Nextt, the acoustic models were separately built with each of the speech databases. The recognition performance was experimentally evalauted using the acoustic models and the relations of the differences of the performance to those of the acoustic features observed in the analyses were investigated quantitatively. The effecitveness of speaker adaptation was also investigated in the same manner.