NN speech recognition utilizing aligned DTW local distance scores

Abstract

This paper presents the neural network (NN) speech recognition using processed LPC input features. But NN has a limitation that the network must have a fixed amount of input nodes. The input feature processing method will use frame matching based on Dynamic Time Warping (DTW) algorithm to fix the input size to a fix amount of input vectors. The LPC features are aligned between the input frames (test set) to the reference (training set) using our DTW fixing frame (DTW-FF) algorithm. This proper time normalization is needed since NN is designed to compare data of the same length, whilst same speech can varies in their length. By doing frame fixing or also known as time normalization, the test set and the training set frames are adjusted so that both sets will have the same number of frames according to the reference set. The neural network with backpropagation algorithm is used as the recognition engine at the back-end processing to enhance the recognition performance. The results compare DTW with LPC coefficients to back-propagation NN with LPC coefficients adjusted using DTW.