Total variation manifold learning method for non-stationary speech features extraction

Abstract The present paper proposes a novel total variational regularized manifold learning method (TV-MLR) for extracting speech features, aiming to enhance the classification performance of speech recognition (SR). This method explores the local linear embedding (LLE) manifold learning approach with total variation constraint. We apply the first and second Fourier transforms to the speech signals for analysing both short-time and long-time features associated with signal amplitude and frequency. These feature vectors are utilized for constructing a high-dimensional (HD) covariance matrix. To reduce the dimensionality of HD matrix, we employ a variational operator on adjacent features to obtain a weight matrix based on total variation, which is further used to optimize k -neighborhoods. By minimizing weight energy, a manifold learning model is construct that effectively clusters and extracts relevant features. Finally, we evaluate our proposed method by examining its impact on SR using support vector machine techniques. Finally, we evaluate our proposed method by examining its impact on SR using support vector machine techniques; experimental results show that the data volume of this method is reduced by about 50% compared with state-of-the-art methods.