A nonlinear feature extraction method for phoneme recognition

The choice of the best parametric representation of acoustic signals is determinant in achieving a high level of accuracy in speech recognition applications. Most state of the art speech recognizers, rely on the Mel-Frequency cepstral coefficients (MFCC) as a feature extraction method, however, this method fails to capture nonlinearities related to the modulation patterns occurring in speech signals. In this contribution, we propose a novel, feature extraction method that partially simulates the frequency analysis and nonlinearities occurring in the human auditory system. This is achieved by using a passive Gammachirp filterbank for frequency analysis and the Dyn operator for nonlinear processing of the speech signals. The performance of the algorithm was tested in various noise conditions including white, pink and subway noises at various signal-to-noise ratios (SNRs). Results show that this method achieves a significant improvement over the MFCC.