A comparison StudyofCepstral Analysis with Applications toSpeech Recognition

Three cepstral parametric methods werecom- order toestimate thetransfer function thevocal tract and paredforspeech recognition application: RealCepstrum, theglottal pulse, undertheassumption that pitch canbe Mel-Frequency Cepstrum anda newmethodMaximum modeled asanimpulse train. Likelihood Cepstrum. Thecepstral parameters wereex- tracted fromtraining andtesting setsthat, consisted of TheRCEPSarecalculated using partoftheTI-DIGITdatabase. Theparameter extraction c (bothstationary anddynamics) was performed bythe T {log .F{} (1) HTK engine andMatlabscripts. Training andrecognition whereF isthediscrete Fourier transform andy isthe wereperformed byHTK,using continues density HMMs. Simulations withadditive noise wereperformed andtheirN point observation oftheprocess y(t)Sincey(t)isa results compared. Themaximum-likelihood cepstrum with convolution ofaninput signal x(t) withthesystem h(t), dynamics hasproved tobesuperior totherealcepstrumthen andsignificantly improved therecognition ratetobealmost