Mel-Frequency Cepstral Coefficients as Features for Automatic Speaker Recognition

Automatic speaker recognizer can be based on the use of mel-frequency cepstral coefficients as speaker features. Mel-frequency cepstral coefficients depend on energy inside considered auditory critical bands. These auditory critical bands model masking phenomena. Application of triangular auditory critical bands results in better recognition accuracy with respect to the case when rectangular auditory critical bands are applied. Recognition accuracy when exponential auditory critical bands are applied outperforms recognition accuracy of automatic speaker recognizer when triangular or rectangular auditory critical bands are applied. Application of transformation on elements of speaker model, which target decreasing of difference between testing and training models of the same speaker, can increase recognition accuracy.