Publication Type : Conference Paper
Publisher : APSIPA ASC 2010 - Asia-Pacific Signal and Information Processing Association Annual Summit and Conference
Campus : Coimbatore
School : School of Engineering
Department : Electronics and Communication
Year : 2010
Abstract : pIn this paper, we study how performance of hidden Markov model - neural network (HMM-NN) phone recognizers can be enhanced using probabilistic features, without actually increasing the number of nodes in the neural network. This is necessary when the amount of labeled data available for training the models is small. We conduct two studies. One is to explore a multilingual probabilistic feature frontend. Another is to develop a multilingual acoustic model. We got an improvement of 2.87 and 4.75 per cent for Hindi and Tamil absolute phone recognition accuracy, and 3.03 and 7.02 per cent improvement for the multilingual phone recognition system for the respective languages./p