An improved method in speech signal input representation based on DTW technique for NN speech recognition system

Abstract

A pre-processing of linear predictive coefficient (LPC) features for preparation of reliable reference templates for the set of words to be recognized using the artificial neural network is presented in this paper. The paper also proposes the use of pitch feature derived from the recorded speech data as another input feature. The Dynamic Time Warping algorithm (DTW) is the back–bone of the newly developed algorithm called DTW fixing frame algorithm (DTW–FF) which is designed to perform template matching for the input preprocessing. The purpose of the new algorithm is to align the input frames in the test set to the template frames in the reference set. This frame normalization is required since NN is designed to compare data of the same length, however same speech varies in their length most of the time. By doing frame fixing, the input frames and the reference frames are adjusted to the same number of frames according to the reference frames. Another task of the study is to extract pitch features using the Harmonic Filter algorithm. After pitch extraction and linear predictive coefficient (LPC) features fixed to a desired number of frames, speech recognition using neural network can be performed and results showed a very promising solution. Result showed that as high as 98% recognition can be achieved using combination of two features mentioned above. At the end of the paper, a convergence comparison between conjugate gradient descent (CGD), Quasi–Newton, and steepest gradient descent (SGD) search direction is performed and results show that the CGD outperformed the Newton and SGD.