A Practical Speaker Independent Voice Recognition Using Segmental Features

概要

論文の詳細を見る
It is well known that segmental features give good performance to voice recognition. However this strategy increases amount of calculations for likelihood estimation in an acoustic model. In order to reduce amount of calculations two strategies are taken here. First one is reduction of segment frames. Second one is using a linear function for likelihood calculation which is obtained from common covariance assumption. Word-spotting technique is also used for keeping robustness of the system. By considering posteriori-probabilities, good performance is obtained when executing word-spotting. We confirmed efficiency of the method by experiments using voice data in car noise environment.
社団法人電子情報通信学会の論文
2002-03-01