A comparison Study of Cepstral Analysis with Applications to Speech Recognition

Three cepstral parametric methods were compared for speech recognition application: Real Cepstrum, Mel-Frequency Cepstrum and a new method Maximum Likelihood Cepstrum. The cepstral parameters were extracted from training and testing sets that, consisted of part of the TI-DIGIT database. The parameter extraction (both stationary and dynamics) was performed by the HTK engine and Matlab scripts. Training and recognition were performed by HTK, using continues density HMMs. Simulations with additive noise were performed and their results compared. The maximum-likelihood cepstrum with dynamics has proved to be superior to the real cepstrum and significantly improved the recognition rate to be almost as high as of the Mel-frequency cepstrum.