Improving children's mismatched ASR through adaptive pitch compensation
No Thumbnail Available
With the progress made in the speech processing over the last few decades, an increasing number of user applications employing automatic speech recognition (ASR) systems are being developed. In such human-machine interaction (HMI) applications, the ASR system is often accessed by both the adults and the children. It is well known that the ASR systems trained on the adults' speech exhibit a severely degraded recognition performance when used for transcribing the speech data from the child speakers and vice-versa. One of the ways to achieve good ASR performance for both the adults and the children is to pool a large amount of data from both the group of speakers in the training of the system. The scarcity of the children's speech corpus makes this approach infeasible. On the other hand, pooling a limited amount of children's data with the adults' training set is not found to be very effective.
Supervisor: Rohit Sinha
ELECTRONICS AND ELECTRICAL ENGINEERING