Application of Tucker Decomposition in Speech Signal Feature Extraction

Speech signal feature extraction is an important part of speech recognition system. We present Tucker decomposition to extract speech feature. Firstly, the preprocessed speech signal is decomposed via three-level Wavelet transform, and the information in different scales is obtained. Next, the conventional feature parameters are extracted from the different scales, and a 3-order speech tensor (frames, scales, feature parameters) could be created. Then, the tensor is decomposed by Tucker decomposition, and projection matrices in different mode are obtained. Thirdly, matrix product is performed between speech tensor and projection matrices in each mode, and mapped results are metricized. Finally, feature system in high order space is built, in other words, speech feature matrices are obtained. The feature system can fully express speech signal features. These matrices can be used for model training and speech recognition. Numerical experiments support the advantage of Tucker decomposition over conventional methods for speech signal feature extraction, furthermore, it is robust to noisy speech.