
COMPUTER ENGINEERING
1999 25 8 Vol.25 No.8 1999



ƵϢ
壬
ժҪع˹еƵϢ˳ƵݴʶͻݵƵ˻ݵƵһ㷽ָӦоеĹؼ⡣
ؼʣƵϢʶ𣻻ݼ
Audio Information Retrieval
Li Hengfeng,Li Guohui
(Dept.of System Engineering,National University of Defense Technology,Changsha 410073)
AbstractThis paper overviewes the main stream of abroad audio ingormation retrieval method,analyzes the general audio data transact technology,such as speech recognition and content-based audio retrieval,brings forward the general method of content-based audio retrieval,and puts out some key problem of audio research.
Key wordsAudio;Information retrieval;Speech recognition;Content based retrieval
ƵϢֻϢҪаýݣƵͼƵѾͨһ ( ֡¡۸ ) мͼƵԲɫӾͬƵҪͨмıĳϢѾ޷ԴƵݵļҪƵϢѾΪϢоص֮һ 
ýǳӾýҪý壬ռϢ 20 ңý塣ֻõĽΪƵ 
ʵһҲʾƵʡλԡܸоý壬ԣ໥ġ벨йأǿƵʡϡгṹԡǵĸойأ˶죬ǿȡɫгԡ 
1 ƵԤ 
ƵбһݣһҲǰҪԤýתȡƵı 
1.1 ʶ 
һһӦģֻԲֲͬͬʶ֣ͿԽڳϢм߽˻ʶĳԣʶ (Automatic Speech Recognition-ASR) Ƶصо򡣶ڻݵƵϢӦȡݵƵƵƥ䣬ӶƵݵķͼ 
ʶѾʵ߳ΪһֱȽʵõļ IBM  ViaVoice  Lernout&Hauspie  Voice Express ֱȽϳɹϵͳ ViaVoice Ϊڼı׼á 
ʵʵõ ASR ϵͳǻ˷ģ (HMM) ġ˷ģ¼絥ʵͳƱ֣ģͲɴʶѵɡ HMM ѵкܶЧ㷨ҳܵģСַڴʻʶϵͳҶԹؼʶϵͳҲЧ 
(1) ؼʶ 
ؼʶָڸƵвѯضĵʻͨҪĹؼʺģͽ HMM ѵʹÿһƥ䡣ڴʻʶϵͳؼʶϵͳȾȷСҶʵʵнϴĵԡڹؼʶȽϵ͵Ĳ SWITCHBOARD ղȻɵĵ绰ԻڶԻǹĳһģЩоͼԶĳһԻ⡣ BBN оԱۺϴʻʶؼʶԽȷ⡣ 
(2) ʻʶ 
ͬڹؼʶ𣬴ʻʶ𽫴תıʽȻеʽ HMM ģ͵ĴʻʶһЩȱ㣺ʲʵУ޷ʶҪģͣҪдıѵмڴˣһ " ӵ " ʷֶΣǶǧƵĵʽ HMM ģֻ͡Ҫõٸڵӵģ͡ӵʺһͿɹĵʡ磬 "right" Ϳ 3 ӵ "R AY T" Ӷɡ⣬ԲͬԵص㣬Ҫоʵĳָʡ ASR һŵǣҪƵ֪ģʿ߲Ȼ ASR ϵͳһܴȱ׼ȷԽϵ͡ض򣬼ʹõʶϵͳҲֻܴﵽ 90 ҵ׼ȷʡʵ绰ԻŹ㲥ֻ 50  60 ׼ȷʡ 
1.2 ˵߼ 
˵߼ƵҪ򣬿ԽݵĶƵľࡣʶԣ˵ǱȽϼ򵥶ʵõļ˵ʲôֻע˭˵ġӦ˵ʶжýķֶһǰ;ں뼶㹻׼ȷĻͿƵƵȶýм˵ߵĸı䡣λѧ Wyse һֻڵƵײķ÷ȽռĲͬһ˵ڲĸı䡣˲ʵֵʱ϶µ˵߻仯˷ƵϣѸЧشƵƵȡϢ 
1.3 Ƶȡ 
ڽƵ֮ǰҪȡƵƵصԡƵȡַһȡߣȣǼǸԻԣƵϵԤϵȡƵУȶƵݽмӴӴС 10  30ms ңȻԼӴݼÿһ֡ɢҶ任 (DFT) ʵϳÿٸҶ任 (FFT) Ӧòͬ㷨Ӧܼ 
ȣǽϳõĸӦʱУһǶÿ֡ȡƽͣȻƽ ( ˷˶ƵӦ޹ )  
ƵйصĸԡԴкܶЧ㷨Ͳ׸ˡ 
ʣȡֵвͬʱ "  " 󡣵λʱĴΪ "  "  Ӧü㣬ʶ档ʸߵζӦʱԽϸߣʵ͵ζӦɼҪ־ 
Mel- Ƶϵ (Mel-scaled frequency Cepstral coefficients-MFCCs) Ƶݾ Z 任ͶóĽһÿ֡ȡ 12 ϵԺܺõرÿ֡䴦£ 
Ԥϵ (Line prediction coefficients-LPC)  LPC ƵѹƵϢ㷽мӦáʵкܶԤⷽԼԤ׼ȷȺЧʡ 

ͼ1 Mel-Ƶϵ
2 ݵƵ 
Ȼ㷺ȵȣҪǧƵƵȡϢĻıвͨģҪµļͼҪȡɫ ƵҪȡؼֻ֡ͬдӹ㷺ƵȡϢܶԲͬƵݽзͼҪõݼ (Content Based Retrieval-CBR) ķ 
2.1 ƵϵͳĻṹ 
Ƶϵͳṹͼ 2 ʾ 

ͼ2 Ƶϵͳṹʾͼ
ͼ 2 ʾƵһȽݿ⣺ƵݽȡƵװݿԭʼƵⲿ֣װⲿ֣ͨƵݾ࣬Ϣװⲿ֡ݿ⽨ԺͿԽƵϢƵҪʾѯ (Query by Example-QBE) ʽûͨѯȷ趨ֵȻύѯϵͳȡֵȷѯʸʸƥ䣬ԴӴС˳ԭʼƵһӦݣͨѯӿڷظûУԭʼƵŵƵݣƵݵ¼ţǶƵоõĲʸռ뱾ֵϢ 
2.2 Ƶ 
ƵӦУҪƵݷࡣƵ;ͬĸָռķָƵҪֳɸࣻжһڵƵһԼķƵľ롣ɲŷ¾룬ԳأӦľ붨塣 
Ƚϳľ෽ǲƽʸ Muscle Fish ˾ӦôַȽʶݼӴÿ֡ȡߣȣȣԣмֵֵ 13 ƽʸķ䵽ʸռضСҪȷĳݵ𣬼뱾ľ룬С뱾ڵ༴Ϊ󡣼һȡʸռ N 㡣 
һֻͽṹķԴʶݼӴû໥Ϣ (Maximum Mutual Information-MMI) ķʸռֳ L ཻȻݸеķֲģ塣ģľ뼴ɽзͼ 
2.3 ѯӿ 
Ƶмжּʽ 
(1) Լ ͨıͬļСʱһԣȡʣƵԡ 
(2) ֵ ǽϸ߲εļĳֵƵݡ 
(3) ʾ (QBE)  ߲εļҲһ "  " ݣ "  " ƵƵݡͼ 3 һȽһ㻯Ĳѯӿڡ 

ͼ3 Ƶ
2.4 ݿ֯ 
ƵݿۺϹϵݿͶýݿŵ㣬ڸ߲ķײѺܳĹϵݿ⣬ DB2  Oracle ԽƵݿ֯͸ЧĵײϵܣмӦ ODBC ӿڣһϵͳ 
ݿ¼Ҫݣ 
һԣļļʽСڵȣ 
Ƶԣȡʣʱȣ 
ԣǿ MFCCs ϵȡ 
ʵӦʱ۴߼ЧʡƵȽ϶࣬ B* ṹ 
3  
ҪعеƵϢ˹Ƶоص㣬˳ƵݴʶͻݵƵ˻ݵƵһ㷽ָӦоеĹؼ⡣ 
Ƶо 20 ꣬ʵõ滹ʱա AOD  VOD Ƶϵͳʵƹ㣬ҪЧƵϢһҪо Web ϵƵ档ѾӴ˺ܶǿ Web 棬 Yahoo, AltaVista, ȽʵõƵǷëǡ Internet ϵĶýǳ޴ҪһЩЧӺ̺ҳҪϢ⣬ʶٶȺ׼ȷҪߣƵķƵݵķҪоĿ⡣ 
ߵλƼѧϵͳѧϵɳ 410073
ο 
1 壬 . ýϵͳ . ʵ ,1997-08 
2 Rabiner L R.A Tutorial on Markov Models and Selected Applica- tions in Speech Recognition. Proc IEEE,1989,72(2)  257-286 
3 Wold E, Blum T, Keslar D. Content-based Classification, Search, and Retrieval of Audio. IEEE Multimedia,fall, 1996  27-36 
4 Wicox L, Chen F. Segmentation of Speech Using Speaker Identi- fication, Proc ICASSP'94 
