DFT based feature extraction with non-uniform spectral compression for robust speech recognition

A DFT based feature extraction method with the use of non-uniform spectral compression (NSC) is proposed. In this method, it generalizes the conventional DFT based feature extraction method and allows different DFT components of each segmented speech having different degrees of compression. It is proposed that the low frequency components are set to have smaller compression than high frequency components, as the low frequency components always have stronger spectral energy, which are good in noise tolerance, than high frequency component according to the heuristic information of speech. By experimental results, it is shown that the proposed method can provide significant improvement in recognition performance over the conventional counterparts under clean and white noise environments.