Speaker adaptation of continuous density HMMs using multivariate linear regression

1 2 1 1 n j j j n H À @ A AE @ A j j j j j j j j j j j j j dependent @ƒsA models —nd —d—pts the model p—r—mE eters to the new spe—ker ˜y tr—nsforming the me—n p—r—meters of the models with — set of line—r tr—nsE formsF „he tr—nsform—tions —re found using — m—xE imum likelihood ™riteri— whi™h is implemented in — simil—r f—shion to the st—nd—rd wv tr—ining —lgoE rithms for rwwsF fy using the s—me tr—nsform—E tion —™ross — num˜er of distri˜utions —nd pooling the tr—nsform—tion tr—ining d—t— m—ximum use is m—de of the —d—pt—tion d—t—F „his —llows the p—r—meters of —ll st—te distri˜utions to ˜e —d—ptedF ‚esults —re presented on the IHHH word e‚€e ‚esour™e w—nE —gement ‚wI d—t—˜—se using — ™ontinuous density q—ussi—n mixture rww system with ™rossEword triE phone modelsF i—™h st—te in — ™ontinuous density q—ussi—n mixture rww h—s —n output distri˜ution m—de up of — numE ˜er of mixture ™omponent densitiesF e st—te with mixture ™omponents ™—n ˜e exp—nded to p—r—lE lel single mixture ™omponent st—tesF „hus the ™—se of single mixture ™omponent st—tes is des™ri˜edD —nd the extension to multiple mixture ™omponents is str—ightE forw—rdF „he pro˜—˜ility density of st—te gener—ting — spee™h o˜serv—tion ve™tor of dimension is @ A a I @P A AE @IA where —nd AE —re the me—n —nd ™ov—ri—n™e reE spe™tively of the output distri˜ution of st—te F „he —d—pt—tion pro™edure is ˜—sed on reEestim—ting the me—ns of the st—te distri˜utions using — line—r tr—nsE form of the existing me—nF „hus it is —ssumed th—t in —d—pting from the ƒs system to the spe—ker —d—pted @ƒeA system the st—te tr—nsition pro˜—˜ilities —nd the ™ov—ri—n™es of the st—te distri˜utions do not ™h—ngeF „he ƒs me—ns —re m—pped to the unknown ƒh me—ns @ A ˜y — line—r regression tr—nsform estim—ted from the —d—pt—tion d—t—X a where is the @ C IA tr—nsform—tion m—trix —nd is the extended me—n ve™torX a ‘I “ sf —n individu—l regression m—trix is used for e—™h st—te using sm—ll —mounts of —d—pt—tion d—t— will reE sult in very poor estim—tes of the m—tri™esF „husD e—™h regression m—trix is —sso™i—ted with m—ny st—te distri˜utions —nd estim—ted from the ™om˜ined d—t—F „ying tr—nsform m—tri™es in this m—nner is simil—r in essen™e to the tying of st—tes or mixtures ‘U“ ‘V“ whi™h m m j n …