Recognition of isolated words corrupted by impulsive noise

The paper deals with the problem of speech recognition when some segments of uttered word are missing. We consider the recogniser that operates on the cepstrum coefficients of the speech signal. Commonly, missing cepstrum coefficients are replaced by the coefficients that are extracted from the neighbouring parts of the missing segment. We investigate an increase of the accuracy of the recogniser by the restoration of the missing speech signal segment with the following extraction of possibly precise cepstrum coefficients. We propose to use signal restoration that is based on the interpolation of missing speech signal characteristics. Therefore the period of the fundamental frequency and intensity characteristics are employed. By experimentation, it is shown that the polynomial approximation of speech signal characteristics improves recognition accuracy.