The use of noisy frame elimination and frequency spectrum magnitude reduction in noise robust speech recognition

In this paper the procedure for feature vector extraction and the problems, which must be solved, by defining the feature vectors, which contain only the information about the speech signal are described. A new procedure of feature extraction which is based on the frame elimination and frequency spectrum reduction for the noisy part of the speech signal is proposed. For all tests the Slovenian telephone speech database SpeechDat II was used. The connected digits were used for both, training and testing. There were 800 speakers used for training and 200 for testing. The word recognition accuracy was increased for 3.1 percentage points with the new procedure, and this was achieved, when the number of Gaussian mixtures was four times smaller than with the ordinary method. The results obtained are especially encouraging for the systems where the size of the available memory and processing power are limited (for example, mobile phones).