Noise robust speech parameterization using multiresolution feature extraction

In this paper, we present a multiresolution-based feature extraction techni que for speech recognition in adverse conditions. The proposed front-end al gorithm uses mel cepstrum-based feature computation of subbands in order no t to spread noise distortions over the entire feature space. Conventional f ull-band features are also augmented to the final feature vector which is f ed to the recognition unit. Other novel features of the proposed front-end algorithm include emphasis of long-term spectral information combined with cepstral domain feature vector normalization and the use of the PCA transfo rm, instead of DCT, to provide the final cepstral parameters. The proposed algorithm was experimentally evaluated in a connected digit recognition tas k under various noise conditions. The results obtained show that the new fe ature extraction algorithm improves word recognition accuracy by 41% when c ompared to the performance of mel cepstrum front-end. A substantial increas e in recognition accuracy was observed in all tested noise environments at all different SNRs. The good performance of multiresolution front-end is no t only due to the higher feature vector dimension, but the proposed algorit hm clearly outperformed the mel cepstral front-end when the same number of HMM parameters were used in both systems. We also propose methods to reduce the computational complexity of the multiresolution front-end-based speech recognition system. Experimental results indicate the viability of the pro posed techniques.