In this paper, we propose new speech features using independent component analysis to human speeches. When independent component analysis is applied to speech signals for efficient encoding, the adapted basis functions resemble Gabor-like features. Trained basis vectors have some redundancies, so we need to select some of the basis vectors by some reordering methods. The basis vectors are almost ordered from low frequency basis vector to high frequency basis vector. This is compatible with the fact that speech signals have relatively more information on low frequency range. Then only a few active coefficients of the trained basis vectors are sufficient for encoding the speech signals. Those trained speech features can be used in automatic speech recognition systems, and the proposed method gives better recognition rates than conventional mel-frequency cepstral coefficients (MFCCs) features. Trained basis vectors can be also applied for the removal of Gaussian noise. Speech signal corrupted by additive white Gaussian noise is almost recovered like clean speech signal after the denoising process. Then, these denoised speech features show better recognition performances than MFCCs features.
