Sil Aa Ae Ah Ao Aw Ax Ay B Ch D Dd Dh Dx Eh En Er Ey F G Hh Ih Ix Favoring Delta Coefficients Favoring Spectral Coefficients 4. Future Work

Many speech recogni ti on systems [Lee88] , [ Shi 85] , [Hua92] , use mul ti pl e i nformationstreams to compute HMMoutput probabi l i ti es (e. g. systems based on semi conti nuous or di screte HMM' s use one codebook for cepstral coe ci ents, andanother one for del ta cepstral coe ci ents). The nal score a wei ghted sumof the contri buti ons of every am. These wei ghts can be found empi ri cal l y l l y the same set of wei ghts i s used for evi c model . There i s reason to bel i eve that eatures whi ch are more important for model s than for others. Especi al l y, the begi nni ng and ending sege more context dependent that case the probah recogni zer shoul d a-spectrumthan 85] , [Boc93] , coe ci ents ndmore . 1. TRAINING whi ch uses n i nformati on t , and for a gi venHMMtreami wi l l compute Pi(xtj S), where xt . The overal l probi s then P (xtj S) = s awei ghti ng factor for al l ed codebook expo1 and P i =1, obabi l i ty. We HMMstate