A Novel Modeling Method for Acoustic Model in Deep Neural Network by Introducing Language Vector

Abstract

Much attention has been paid on Multi-language or cross-language speech recognition domain. A kind of speech recognition system with one language and different accents is also an application in the above domain. In this paper, a new concept called language vector is proposed and then is applied in the learning of acoustic model in deep neural network. The proposed method introduces language vector by conditional learning and multi-task learning and dramatically improves the performance of the English speech recognition system aimed at British accent and Chinglish accent.