Performance Study of TDNN Training Algorithm for Speech Recognition

Time delay neural network is special types of architecture of artificial neural network whose functionality is to work with continuous data. It permits a classification of the phoneme by taking into account the dynamic aspect of speech and consequently to overcome problems of co articulation phenomenon. This paper concern about the performance of training algorithm used to train TDNN. Set of speech data are used to analyze the performance of training algorithm. The feature of speech data are extracted from speech spectrogram by using Otsu’s thresholding method. The extracted features are fed into TDNN with 8 different training algorithms to adjust the synaptic weights in order to minimize the error between the computed output and the desired output for all samples. Fifty percent of speech data are used to training and other fifty percent of data are used to test the network. From the experiment It have been seen that scaled conjugate gradient outperform over other algorithm. Conjugate Fletcher-Reeves Update Achieved average accuracy 95% for unknown and 99% of known speech word.