Temporal Speech Normalization Methods Comparison in Speech Recognition Using Neural Network

摘要

Speech signal is temporally and acoustically varies. Recognition of speech by static input Neural Network requires temporal normalization of the speech to be equal to the number of input nodes of the NN while maintaining the properties of the speech. This paper compares three methods for speech temporal normalization namely the linear, extended linear and zero padded normalizations on isolated speech using different sets of learning parameters on multi layer perceptron neural network with adaptive learning. Although, previous work shows that linear normalization able to give high accuracy up to 95% on similar problem, the result in this experiment shows the opposite. The experimental result shows that zero padded normalization outperformed the two linear normalization methods using all the parameter sets tested. The highest recognition rate using zero padded normalization is 99% while linear and extended linear normalizations give only 74% and 76% respectively. This paper end before conclusion by comparing data used from previous work using linear normalization which gave high accuracy and the data used in this experiment which perform poorer.