O-Engineers O-Engineers Aug 2017 | Page 45

For text-dependent speaker recognition, improved performance has been found by introducing ceps- tral derivatives into the feature space as it captures the transitional information in the speech. Cepstral weighting or liftering is also used which enhances the speaker recognition. Cepstral mean subtrac- tion (CMS) technique also significantly improves the performance of a recognition system in which training is done on one channel condition while testing is done on another channel condition. An- other technique, known as Pole-filtered Cepstral Mean Subtraction (PFCMS), modifies the LP poles so as to broaden the bandwidth of the formant poles. The cepstrum formed from these modified poles has less speech information and more chan- nel information. It has noticed that PFCMS outper- forms CMS in speaker-identification problem. The relative spectral (RASTA) technique takes ad- vantage of the fact that the rate of change of nonlin- guistic components in speech often lies outside the typical rate of change of the vocal-tract shape. 45