A Speech Recognition System Based Improved Algorithm of Dual-template HMM

Transcription

1 Avibe onine Procedi Engineering 5 (20 ) A Speech Recogniion Sysem Bsed Improved Agorihm of Du-empe HMM JingZhng,b Min Zhng c * Fcuy of Auomion, Gungdong Universiy of echnoogy, Gungzhou Gungdong 50006, Chin b Dep. of Compuer Science nd Technoogy, Gungdong Universiy of Foreign Sudies, Gungzhou Gungdong 50006, Chin c Fcuy of informion, Gungdong Universiy of echnoogy, Gungzhou Gungdong 50006, Chin Absrc The hidden Mrkov (HMM) nd speech recogniion gorihm bsed his mode were sudied in he pper. In ddiion he mode nd recogniion gorihm of HMM go be improved bsed on he rdiion he HMM. In he process of modeing, hrough he rining of muipe observe sequence o chieve he recogniion of non-specific peope, nd ccording o he differen number of HMM ses o esbish he doube-empe of rough nd high precision, nd hrough he second mching gorihm o chieve higher recogniion re. A speech recogniion sysem combined MFCC prmeers nd HMM gorihm ws consruced bsed improved HMM gorihm. Experimen resu shown he speech recogniion re of rge vocbury of non-specific peope ws grey improved. 20 Pubished by Esevier Ld. Seecion nd/or peer-review under responsibiiy of [CEIS 20] pen ccess under CC BY-NC-ND icense. Keywords:Speech Recogniion; Hidden Mrkov Mode; Doube Tempe Mching; Muipe observe sequence; D overfow. Inroducion Speech sign is smooh nd ime-invrin in very shor nysis inerv, nd he sisic feure cn be described by he prmeers of cssic iner mode, bu for n over speech sign, he sign is ime-vrying nd he prmeers of iner mode so woud chnge. As sisic mode, HMM coud boh represen he informion of shor-erm smooh iner mode, nd describe he rnsiion beween ech mode. Therefore, HMM ws used o buid sisic modes for he speech sign: one is Mrkov chin wih finie number of ses used o simue he impied sochsic process of he speech sign wih chnging sisic properies, nd he oher is he sochsic process of observion sequence ssocied wih ech se of Mrkov chin, s n ide souion of speech recogniion, he HMM compeey expressed he cousic mode of speech. However, for he prcic ppicion of HMM in speech recogniion process here re si mny probems, such s he rge number of observion sequences invoved in he rining, he esime of inii prmeers, nd so on, he key of improving speech recogniion re is how o resove hese issues. * Corresponding uhor. Te.: E-mi ddress: h_go@63.com Pubished by Esevier Ld. doi:0.06/j.proeng pen ccess under CC BY-NC-ND icense.

2 Jing Zhng nd Min Zhng / Procedi Engineering 5 ( 20 ) The improved HMM mode nd recogniion gorihm 2.. The improved HMM modes ---- muipe observions sequence modeing For he cssic Bum-Wech gorihm, he prmeer revuion formu ws deduced under he condiion of ssuming ony one observion sequence. In he ppicion, here re rge number of observion sequence invoved in rining, h is, for ech HMM mode, rge number of speech d wi be coeced, he respecive sequence of MFCC prmeers shoud be ccued, nd hen used for he prmeers revuion of corresponding HMM. For exmpe, for he buiding HMM modes of word "b", i shoud find o of peope, nd record muipe wv fies for "b" of everybody, fer he endpoin deecion, hen o ccue prmeers sequence of MFCC, h is so-ced he observion sequence, hen he Prmeers of he mode cn be rined. In cu ppicion, usuy more hn one observion sequence were used o rin HMM, hen when rin HMM wih L observions sequence, he revuion formu of Bum-Wech gorihm ( ) shoud o be mended. Assuming L observions sequence were, =,2, L,, in which =,,, nd ssuming ech observion sequence ws independen, hen go formu (): T L = 2 = Since he revuion formu ws bsed on he frequency of differen ime, herefore, foe L rining sequences, revised revuion shown s he formu (2),(3),(4). L i i = ( ) β ( ) π, i N = () (2) ij = ij j = = = = ( i) b ( + ( i) β ( i) ) β +, i, j N (3) b jk = = = nd = v = = k β β, j N, k M (4) 2.2. The improved recogniion gorihm of HMM-----doube-empe mching The mos imporn issue of HMM mode rining is he esimes of inii prmeers. Differen inii rining my produce differen resus nd he pproprieness of inii esimes is so reed o he fin mode prmeers cn wheher converge o he gob opimum or no. The improvemen for rdiion HMM recogniion gorihm is miny doping wo-empe mching mehod for differen empes iniiizion. For speech mode, he prmeers of coninuous HMM re more hn h of discree HMM, nd i coud chrcerize he spi disribuion of feure vecor more ccurey. Bu is probems re rge ccuion nd he convergence is sow.

3 2288 Jing Zhng nd Min Zhng / Procedi Engineering 5 ( 20 ) ne of he prominen performnce is, in heory, he mode cn be chieved convergence fer finie ierions of ny inii vue. In fc, if improper inii vue were chosen, combined wih he imprecise chrcerizion of mode prmeers, which my cuse ierion divergence or excessive ierions nd resu in ong rining ime or rining resus do no converge, hen he requiremen of cu ppicion coudn be chieved. I ws found in he reserch nd experimen h he inii vue seecion of mode prmeers nd A hd ie effec, in gener, i ws seeced rndomy or meny. In he rdiion HMM mode he iniiizion of he B is genery ken o men. Th o divide observion vecor sequences of ech primiive o be idenified ino N segmens (N is he number of ses.). Since HMM mode hs he srucure from ef o righ, he observion vecor is corresponding o ech se in ime. Afer Segmens dividing, ech observion vecor corresponds o se of HMM, nd hen o ccue he men correion, vrince nd oher prmeers of ech secion. Bu he inii processing of B ccording o his pproch hs cerin reionship wih he seecion of he ses quniy, h he fewer ses, he rough division, nd he chrcerisics of he speech sign coud no be we refeced. The speech recogniion is miny pern mching hrough ech HMM empe in sysem, h he pern mching by Vierbi gorihm, hen seec he speech h ws coses o he speech o be recognized s he recogniion resus ccording o mching probbiiy. Therefore, in he impemenion process of sysem he mui-empe mching gorihm ws proposed, h is, hrough wo mches o chieve speech recogniion nd he improvemen of recogniion ccurcy. The bsic ide of he wice mching gorihms is: he sysem rined wice, nd he rining ws divided ino he HMM empe of rough ccurcy nd high precision, he number of ses N in he HMM of high precision is rger hn h of rough ccurcy. In he recogniion process, firsy o mch he speech o be recognized wih he HMM prmeers of rough ccurcy, nd hen ccording o he recogniion resu, ke he N recogniion resus wih high simiriy probbiiy s he bsis of he empe need o be mched when seecing he high ccurcy empe, nd hen o he second mch wih he empe wih high ccurcy, hereby o improve speech recogniion ccurcy. In heory, he rger number se is he beer becuse h he error re of recogniion wi be reduced o sbe eve wih he increse in he number of ses. However, since he rining smpes is imied, herefore, he number of ses N cn no be oo rge, oherwise, he rining mny of he corresponding iems of ses in he prmeer = (A, B, ) woud be 0 or very cose o 0 s redundncy, he experimen number of speech ses fixed rnging from 3 o 8 ccording o he compexiy. Tken Engish number 0-9 voice s es d, experimens were crried wih differen number of ses, Tbe shown he verge experimen resus of hree imes recogniion wih he empes of differen ses number. I cn be drwn from he d in Tbe h wih he incresing number of ses N, he sysem recogniion re incresed ccordingy, bu s he sme ime, he recogniion ime sysem required so incresed, which is due o he ccuion of he sysem incresed, h he compexiy of he sysem incresed. Tbe Recogniion resus of he empes wih differen ses number

4 Jing Zhng nd Min Zhng / Procedi Engineering 5 ( 20 ) Number of ses Number of empes Number of verge recogniion Averge recogniion re % % % % % % Averge recogniion ime 3. Sysem evuion The recogniion modue of sysem ws crried by using du-empe mching. In he recogniion process, in order o improve he recogniion efficiency, he wo empes ws no used for ech recogniion, if he firs idenificion hs been successfu, empe of rough ccurcy used ony, ony he firs unsuccessfu recogniion or he recogniion resu ws no mong he op wo resus hen he second recogniion ("re- recogniion "or "wice recogniion ")used. As Fig. shown, i is successfu fer wice recogniion. Fig.. resu of wice recogniion The mui-empe mching of four cegories of spors vocbury in he rining window were crried, nd he experimen d s shown in Tbe 2 Tbe 2 he recogniion resu of du-empe mching cegory The verge recogniion re of once recogniion The verge recogniion ime of once recogniion The verge recogniion re of re-recogniion Bskeb css 90.2% % Swimming css 89.76% % Trck nd Fied css 88.57% % The verge recogniion ime of rerecogniion From he d in Tbe 3 cn be seen, fer using du-empe, he recogniion re of sysem hs been incresed, bu i is sme o he recogniion ime, he over verge recogniion re of he sysem reched 90% or more, which mee he requiremens of Sysem performnce.

5 2290 Jing Zhng nd Min Zhng / Procedi Engineering 5 ( 20 ) Concusion A du-empe mching mehod ws proposed fer he improvemen of HMM gorihm imed he rge vocbury of non-specific, nd some effecive souions bou such issues s he inii mode seecion of HMM gorihm encounered in he prcic ppicion of, muipe observions sequence invoved in he rining nd D underfow, which mde he speech recogniion re of rge vocbury of non-specific chieved more hn 90% in he PC pform. However, hough he sysem using HMM gorihm hs some dvnges, bu here re so some disdvnges, such s wih he number of recogniion empe incresed, he o ses umber of HMM mode so incresed exponeniy, which wi require rge sorge spce. In order o ge robus nd cn we ppied o he mode of embedded sysems; some se cusering mehod need be sudied. Acknowedgemens This work is priy suppored by The minisry of educion of humniies nd soci science projec #0YJCZH220. Reference []FengQin Yng,Chnghi Zhng, GeBi. A nove Geneic Agorihm Bsed on Tbu Serch for HMM pimizion. Fourh Inernion Conference on Nur Compuion.Jinn,2008;p.57-6 [2]Zho Hui, Gu Y-qing, Tng Cho-jing. Speech Recogniion Mehod of Du-mode Bsed Muipicion HMM, Compuer Engineering,vo. 36, Dec. 200, p. 7-9 [3]Hu Lei, Lu Luo-xin, Hung To. The Appicion of An Improved Hidden Mrkov Mode in Speech Recogniion.Informion nd Conro, vo. 36, June. 2007, p [4]Yu Mei-jun, M Xi-rong. An Improvemen of dynmic gesure recogniion bsed HMM. Compuer Science, vo. 38, Feb. 20, p [5]Liu Xun-he, Song Ting-xin. Speech Recogniion nd Conro Appicions. Beijing: Science Press, 2008,pp [6]Zhng Jin-ping, Li Ming, Suo Hong-bin. The Appicion of Long speech feures in he speker recogniion, Journ of Acousics, vo. 35,Feb. 200,p [7]Zho Hui, Tng Chojing, Yu To. Fs Threshoding Segmenion for Imge wih High Noise. Proc. ICIA 08, Dec.2008, p [8]Liu Qing-sheng, Xu Xio-peng, Hung Wen-ho. A sudy of speech endpoin deecion. Compuer Engineering,,vo. 29, Mr. 2003, p [9]Li Shou. A sudy of Speech Feures Exrcion. Xi'n: Xi'n Universiy of Eecronic Science nd Technoogy press, 2005,p [0]Ye Qing-yun, Jing Ji. An Improvemen Agorihm Bsed MFCC feure of Speech. Wuhn Universiy of Technoogy, vo. 29, My. 2007, p []He Qin, Liu Ji. The pimizion Mehods bou he Se number of he HMM mode in Chinese coninuous speech., Journ of Informion Processing, vo. 20, June. 2006, p