Hybrid Method for Digits Recognition using Fixed-Frame Scores and Derived Pitch

Abstract

This paper presents a procedure of frame normalization based on the traditional dynamic time warping (DTW) using the LPC coefficients. The redefined method is called as the DTW frame-fixing method (DTW-FF), it works by normalizing the word frames of the input against the reference frames. The enthusiasm to this study is due to neural network limitation that entails a fix number of input nodes for when processing multiple inputs in parallel. Due to this problem, this research is initiated to reduce the amount of computation and complexity in a neural network by reducing the number of inputs into the network. In this study, dynamic warping process is used, in which local distance scores of the warping path are fixed and collected so that their scores are of equal number of frames. Also studied in this paper is the consideration of pitch as a contributing feature to the speech recognition. Results showed a good performance and improvement when using pitch along with DTW-FF feature. The convergence rate between using the steepest gradient descent is also compared to another method namely conjugate gradient method. Convergence rate is also improved when conjugate gradient method is introduced in the back-propagation algorithm.