Turkish Journal of Electrical Engineering and Computer Sciences




We propose a new feature extraction algorithm that is robust against noise. Nonlinear filtering and temporal masking are used for the proposed algorithm. Since the current automatic speech recognition systems use invariant-integration and delta-delta techniques for speech feature extraction, the proposed algorithm improves speech recognition accuracy appropriately using a delta-spectral feature instead of invariant integration. One of the nonenvironmental factors that reduce recognition accuracy is the vocal tract length (VTL), leading to a mismatch between the training and testing data. We can use the invariant-integration feature idea for decreasing the VTL effects. The aim of this paper is to provide robust features that provide improvements in different noise conditions as well as being robust against VTL effect changes. This results in more improvement of the recognition accuracy in comparison with mel-frequency cepstral coefficients and perceptual linear prediction in the presence of different types of noises and scenarios.


Robust speech recognition, vocal tract length, temporal masking, invariant integration

First Page


Last Page