о뷢չ
JOURNAL OF COMPUTER RESEARCH AND DEVELOPMENT
1999 36 9 Vol.36 No.9 1999



һֻںϵĶ໰Ͳʶ𷽷
Աΰʥ
ժҪʶʺͶԻӦһʶϵͳҪܣʶʵķͨĽģýϸߵʶʣģ͵ĸӻԼģѵ.⣬˵˺˷λò̶£ЩʶЧܲ.һö໰Ͳֱʶһںϼʶдʶ𷽷.ʵ÷ϵͳԻӦڵģʶʲߵ£Ȼܵõϸߵʶ.
ؼʡʶںϣ໰Ͳ
ͼšTP391.4
A METHOD OF MULTI-MICROPHONE SPEECH RECOGNITION BASED ON DATA FUSION
ZHAO Yi-Bao, WANG Qi, NIE Wei,and SUN Sheng-He
(Department of Automatic Test and Control, Harbin Institute of Technology, Harbin 150001)
AbstractRecognition accuracy and adaptability is two of the most important capacity of a speech recognition system. A general method to improve speech recognition accuracy often focuses on improving acoustic model to obtain a higher recognition accuracy, which often leads to a complex model and difficult model training. In addition, in the case of unfixed position of speaker to microphone, its performance is often very bad. A new speech recognition method is brought forward,which uses multi-microphone to recognize Chinese syllables separately, and data fusion technique to process the recognition results. The preliminary experiment shows it not only can improve the system adaptability to environment, but also can reach a higher system recognition accuracy with a lower accuracy of each acoustic model. 
Key wordsspeech recognition, data fusion, multi-microphone
1
ʶʺͶԻӦʶϵͳҪ.ָʶ𷽷ʶʵķһ㶼ڸĽģýϸߵʶʣģ͵ĸӻԼģѵ.⣬ʶ𷽷ڰ˵˶˷ľϽλԹ̶µʶߵо.Щڻӡ˵˺˷λá벻̶Чܲ1,2.
ںϼǽϢбȽŵļ34ݣָͬһԴĲͬ۲дԻöԱ۲͹ʶһַ.ڹ۲ʱȶַöһϢȻɴжϣӶﵽԸĿ͹ʶ.ںϼʵǶʶһģ.
ǻںģԴϢ˼룬һֻںϵĶ໰Ͳʶ𷽷ͨͲֱɼݣÿһͲֱģֱͣʶȻںϼԸģ͵ʶдԴﵽʶʺͶԻӦĿ.ͻڵһͲʶϵͳȣ÷ڸģʶʲߵ£ԿԻýϸߵʶʵŵ㣬ңڲö໰ͲһԽϺõؿ˷˵زͬɵʶ½⣬⻹˶ԸģʶʵҪ󣬼ģ͵ѵ.ʵ֤˸÷Ч.
2໰Ͳʶԭ
໰Ͳʶԭͼ1ʾ.ݾ˷ɼ͵Եȡȡȡ͵Եģͽʶ𣬵óģ͵ĳʶdiЩʶΪѡںĽںϣںĸһں㷨Ժѡдõʶt.

ͼ1໰Ͳʶԭ
2.1ģ
ĵģͲһָĽĵݹʱǩݹ(TTRNN)Ժڽз.ʱǩݹṹͼ2ʾһһʱӵTTRNN.УTime-TagԪΪʱǩΪÿһ֡u(t)һʱǩ.

ͼ2ʱǩĵݹ
ÿһ֡u(t)Լǰһ֡ķx(t)TTRNNóʱtʱu(t)ڷΪx(t)ʱĳĸPk(u(t))Լʱӷx(t+1).ʱģʽuĳĸΪ
Pk(u)=pk(u(1))pk(u(2))pk(u(n))=pk(u(t))
(1)
ͼ2ڵϵԲȦʾʽ(1)Ĺ.
TTRNNԺڽз࣬ҪһڸĸʣݸʵĲͬȷúһ.ʽ(1)֪ҪPk(u),ֻPk(u(t))ɣTTRNNõ.
2.2ں㷨
ǰ˵ģгʶ𣬵õʶdi,i=1,2,NɺѡD͵ںĽоõʶt.Hj,j=1,2,,mʾm࣬Hj, j=1,,m-1ʾźż࣬mHmʾźż256.
ģͳƶģںںѡDĻϸһں㷨õϵͳʶ
t=f(d1,d2,,dN)
(2)


(3)
ΪHjĸȻȣj=1,,m-1Ȼо׼ںĵоΪ

(4)
ʽ(4)֪ҪøоںϣֻҪLj(D).Lj(D).
S1=idi=j,i=1,,N
S2=idi=m,i=1,,N
S3=idi=l,i=1,,N,ljm
(5)


(6)
PiDjj=P(di=j/Hj)ΪǰģiȷjоĸʣPiMjm=P(di=m/Hj)ΪǰģijоΪźmĸʣPiEjl=P(di=l/Hj)ΪǰģijоΪlĸ.
ͬ

(7)
PiFmj=P(di=j/Hm)ΪǰģiźmоΪjĸ, jS1PiDmm=P(di=m/Hm)Ϊǰģ iȷоΪźmĸʣPiEml=P(di=l/Hm)ΪǰģiźmоΪlĸ, lS3.
ʽ(3)(6)(7)

(8)
PiDjjPiMjmPiEjlPiFmjPiDmmPiEmlͨͳƵõ.
3ʵ鼰
ʵо3Ͳʵֺڵʶ⣬ʵʱ3̨ļͬʱһ˵ķ¼Ͳ˵˵ǰƫƫ3λã¼ʱҪ˵˹̶.ʵ鹲ɼ10Ժ10Ů0910ֵĺΪѵͲݣƵʲóֵ֧11025Hzÿռ16Bit.ݾȡÿһƴݣȻȡ.ȡ¹̽УȶԤأȻÿһƴз֡220Ϊһ֡֡110㣬Ŷÿ֡źżHAMMINGΪ256㣬Ȼõһʸ15 Mel Cepstrum ϵ֡֡ʡ֡ƽȹ20άʸ.ÿһڵгȹõյ.
3Ͳݷֱ3TTRNNģͽѵѵòϷֱԸģͣõʶ.
ڱʵʶ10֣10࣬m=10.ڱʵУǰ0Ϊm࣬Ϊźŷ6ݣʽ(6)ʽ(7)֪PiDjjʵǰ˵iģͰΪjʱоΪjĸʣ N(Ij=j)ΪjʱģʶΪjĸN(Ij)Ϊ뵽ģ͵jܵĸ.ͨͳƵõˣѵõPiDjj.ͬģǿԵõPiMjmPiEjlPiFmjPiDmmPiEmi13ʾ.
 1Ͳ1ʶʼʶ

0123456789
00.80.100000.1000
10100000000
20010000000
30.100.10.50.100.10.100
40.1000.10.500.10.200
50.100000.7000.10.1
60000001000
70000000100
8000.1000000.90
900000000.300.7

2 Ͳ2ʶʼʶ

0123456789
00.8000000.2000
10.20.800000000
20010000000
30.100.20.50.100.1000
40.10000.600.10.200
5000000.80000.2
60000000.900.10
70000000100
8000.1000000.90
900000000.100.9

3 Ͳ3ʶʼʶ

0123456789
00.30.100000.6000
10100000000
20010000000
3000.20.60.100.1000
400000.600.20.200
5000000.80.1000.1
60000001000
70000000100
8000.1000000.90
900000000.200.8

Щݺ󣬼ɸʽ(8)о(4)ʵ3ģʶںϣõʶ.ʵ鹲4T1T2T3T4Խ4ʾ.ڱ4У4ÿһݶÿһģ͵ʶ׼ȷʼ3ģ͵ƽʶʣһǾںϵʶʣӱ4ԿͨںϣԺܺõʶ.
4 ںǰϵͳʶ

T1T2T3T4
ںǰͲ160%60%60%60%
Ͳ250%50%50%80%
Ͳ360%60%60%70%
ƽ56.7%56.7%56.7%70%
ںϺ70%60%80%90%

4
ʶʺͶԻӦʶϵͳûҪܵķѾкܶ࣬ͨĽģýϺõʶʺͶԻӦ.ںϼĻϣһֶ໰ͲʶͨںʵֶͲʶĻϵͳʶ.ʵܹܺõϵͳĻӦڸģʶʲߵ£ԿԵõϸߵϵͳʶ.
ȻӲͶʽϴռý϶ڴʱȱ㣬ĿǰżķչЩȱ㽫ٳΪҪϰ.ü򵥵ķýϸߵʶʺͶԻϺõӦŵ㽫ʹø÷˻ܽӿڡܻˡڶõܺõӦ.
߼飺ԱУ19696ʿоҪоΪʶ硢ȻԴ.
ԱУ1969ʿоҪоΪʶ硢Ȼ.
У1944ڣʿʦҪоΪܲ뼼Ϣ.
ΰУ1960ʿоҪоΪںϣܲ뼼.
ʥͣУ1937ڣʿʦҪоΪơźŴϵͳʶ.
ߵλҵѧԶϵ150001
ο
1Lin Q et al. Microphone array and speaker identification. IEEE Trans. Speech & Audio Proc, 1994, 4(2): 622629
2о. źִ.ӹҵ. 1995
(Yang Xingjun et al. Speech Signal Digit Process(in Chinese). Beijing:Electrical Industry Publisher. 1995)
3Baek W, Bommareddy S. Optimal m-ary data fusion with distributed sensors. IEEE Trans on AES. 1995, 31(3): 11501152
4Chairm Z, Varshney P K. Optimal data fusion in multiple sensor detection system. IEEE Trans on AES. 1986, 22(1):98101
5ΰ. ڼ۵ں㷨о۲ʿģ. ҵѧ, , 1998
(Nie Wei. Research of data fusion algorithms based on hypothesis testing theoryPh D Dissertation(in Chinese). Harbin Institute of Technology, Harbin, 1998)
6η. źż.ҵѧ. 1988
(Duan Fengzeng. Signal Detection Theory(in Chinese). Harbin Institute of Technology Publisher. 1988,60103)
ԭյڣ1998-12-03޸ĸյڣ1999-03-16.
