Speech recognition technique has approached maturity as people spent many years in studying this subject, and it has been employed in our daily life. But disadvantages still exist in the practical applications of speech recognition techniques, especially for recognition of children’s speech. We found that 84% of data recorded by the robot are collected from children. However, recognition experiments using acoustic models trained from adult speech and tested against speech from children show performance degradation clearly. In this paper we focus on the key techniques of children’s speech recognition. There are several aspects in my work: 1.Children’s speech analysis. Based on the children’s speech database, we measured together with the pitch and formant frequencies, analyzed the age effects on children’s speech, and figured out the speech difference between children and adult. 2.Research on children’s speech adaptation techniques. Some recognition experiments have been done using several different acoustics models. One of these models is trained from children’s speech, one is from boys’ speech and another one is from girls’. For improving the performance of children’s speech recognition, a new approach which based on vocal tract length normalization by changing the scale threshold dynamical is introduced. 3.Research on the speech interactive technology and module. This paper introduced embedded system implementation on DSP, optimization techniques, dialogue management and so on. The applications of the module were introduced too.
修改评论