基于动态贝叶斯网络的连续语音识别框架及其Token传递模型.pdfVIP

计算机研究与发展 ISSN N 111777TP J ournal of Comp ut er Research and Develop ment 45 ( 11) : 1882 189 1 , 2008 基于动态贝叶斯网络的连续语音识别框架及其 Token 传递模型 苗夺谦 　王睿智 　冉 　巍 ( 同济大学嵌入式系统与服务计算教育部重点实验室 　上海 　20 1804) ( 同济大学计算机科学与技术系 　上海 　20 1804) (wei_ran @mail . tongj i . edu . cn) A Dyna mic Bayesian Net work Based Frame work f or Continuous Speech Recognition and its Token Passing Model Miao Duo qian , Wan g Ruizhi , and Ran Wei ( M inis t ry of E d ucation Key L aboratory of Embed ded Sy s tems and S ervice Comp uting , Tongj i Univers ity , S hang hai 201804) ( D ep a rtment of Comp ute r S cience an d Technology , Tongj i Uni vers i ty , S hang hai 20 1804) Abstract 　Recent ly , dynamic Baye sian net wor k ( DBN ) ba sed sp eech reco gnition ha s arou sed an increa sin g int ere st , becau se of it s int erp ret abilit y , f actorization and ext en sibilit y , w hich hidden Mar kov mo del s ( HMM s) lack . Alt hou gh a huge succe ss of t he int ro duction of DBN s into sp eech reco gnition in many area s and DBN s ha s been p re sent ed wit h p ro mi sing pot ential to overco me inherent limit ation s of HMM s in sp eech reco gnitio n , p reviou s wor k on DBN ba sed sp eech reco gnition mainly focu se s on i solat ed wor d sp eech reco gnition , and t he f ramewor k s and reco gnition algorit hm s for DBN ba sed continuou s sp eech reco gnition are not a s mat ure and flexible a s t ho se for HMM ba sed o ne . Thi s p ap er i s t rying to addre ss t he p roblem s of flexibilit y an d ext en sibilit y in DBN ba sed continuou s sp eech reco gnition . To achieve t hi s p urpo se , t he to ken p a ssin g mo del , w hich wor k s very well to addre ss t he above p roblem s for HMM ba