A noise robust speech features extraction approach in multidimensional cortical representation using multilinear principal component analysis

摘要

In this paper, we propose a new type of noise robust feature extraction method based on multidimensional perceptual representation of speech in the auditory cortex (AI). Different coded features in different dimensions cause an increase in discrimination power of the system. On the other hand, this representation causes a great increase in the volume of information that produces the curse of dimensionality phenomenon. In this study, we propose a second level feature extraction stage to make the features suitable and noise robust for classification training. In the second level of feature extraction, we target two main concerns: dimensionality reduction and noise robustness using singular value decomposition (SVD) approach. A multilinear principal component analysis framework based on higher-order SVD is proposed to extract the final features in high-dimensional AI output space. The phoneme classification results on different subsets of the phonemes of additive noise contaminated TIMIT database confirmed that the proposed method not only increased the classification rate considerably, but also enhanced the robustness significantly comparing to conventional Mel-frequency cepstral coefficient and cepstral mean normalization features, which were used to train in the same classifier.