Method of speech signal feature extraction based on tensor decomposition model

In this paper a method of speech signal feature extraction was proposed based on tensor decomposition. Speech signal was preprocessed firstly, and the information in different scales was obtained via wavelet decomposition of frame information. Next the conventional feature parameters were extracted from the different scales, and a 3-order tensor (frames scales feature parameters) could be created. Finally projection matrices in different modes were obtained via tensor decomposition, and a feature system in high order space was built. The feature system could fully express speech signal features. The experimental results indicate that compared with conventional feature system, the method proposed is beneficial to the improvement of speech recognition system properties; furthermore it is robust to noisy speech.