NN with DTW-FF Coefficients and Pitch Feature for Speaker Recognition

Abstract

This paper proposes a new method to extract speech features in a warping path using dynamic programming (DP). The new method presented in this paper described how the LPC feature is extracted and those coefficients are normalized against the template pattern according to the selected average number of frames over the samples collected. The idea behind this method is due to neural network (NN) limitation where a fixed amount of input nodes is needed for every input class especially in the application of multiple inputs. The new feature processing used the modified version of traditional DTW called as DTW-FF algorithm to fix the input size so that the source and template frames have equal number of frames. Then the DTW-FF coefficients are retained and later being used as inputs into the MLP neural network training and testing. Thus, the main objective of this research is to find an alternative method to reduce the amount of computation and complexity in a neural network for speaker recognition which can be done by reducing the number of inputs into the network by using warping process, so the local distance scores of the warping path will be utilized instead of the global distance scores. The speaker recognition is performed using the back-propagation neural network (BPNN) algorithm to enhance the recognition performance. The results compare DTW using LPC coefficients to BPNN with DTW-FF coefficients; BPNN with DTW-FF coefficients shows a higher recognition rate than DTW with LPC coefficients. The last task is to introduce another input feature into the neural network, namely pitch. The result for BPNN with DTW-FF plus pitch feature achieved its high recognition rate faster than the combination of BPNN and DTW-FF feature only.