PREDICTORS IN SPEECH PROCESSING
P. Kabal, Ravi P. Rarnachandran · 1988
The statlstlcal correlation between speech samples motivates the use of predlctlon error filters that can remove thls redundancy Formant fllters act to remove the near-sample redundancles while pltch fllters remove the far-sample redundances for samples separated by the pltch period The use of these fllters IS especlally beneficlal In low bit rate speech codmg The coefflclents of these predictors are usually determined sequentlaliy Thls paper discusses Jolntly optlmlzed solutions for the two predlctors The fllters are Implemented ~n transversal form wth the formant fllter always precedtng the pltch fllter (F-P cascade) The Jolnt'ly optlmlzed methods are further subdtvlded Into a comblned solutlon and an lteratlve sequentlal solutlon Both yeld hlgher prediction galns than their conventional sequentlal determined counterpart Furthermore. a practical verslon of the Iterated sequentlal approach In whlch the filters are constralned to be mlnlmum phase generates decoded speech of hlgh perceptual quallty ivhen applled In a codlng environment