The application of time series smoothing function and warping magnitude function to the noisy speech recognition using modification of the spectral shape

摘要

The valleys of spectral envelopes are vanishing, and the spectral value on a low spectral level is rising under the noisy environment and these changes give bad influence to the performance of recognition systems. We proposed the method for the extraction of spectral features of signals using modification of spectral shapes by rules. However, the incorrect valleys are sometimes added in poor environment. In this paper, the weighting functions are defined for the time series of cepstral coefficients obtained from modified spectral envelopes for removing incorrect valleys added by rule. And we also proposed the method for evaluating spectrum on warped magnitude axis. In this method, local peaks at the low frequency area are emphasized. The cepstral parameters applied these methods are used for recognition experiment and performances of parameters are compared.