о뷢չ
JOURNAL OF COMPUTER RESEARCH AND DEVELOPMENT
1999ꡡ3611ڡVol.36No.111999



ʶʶɷģ״̬
״̬פصѧ
ġͤ׿˳
ժҪʶϵͳʵתʵӦãʶͱҪ.ΪʶϵͳʶѡĽ/ʶо⣬˻ɷģ(HMM)ʶϵͳ״̬״̬פصѧ׼.״̬ʸƽ۲ʺ͸ʸ״̬ĺʾȽ趨ͳһľʶޣҲרŵѵ.״̬פֲطǻפֲʺۣ趨һʶޣͬʱɸʶѡ״̬פŶ.ʵʶ׼ܺܺõؾʶʶѡʹʱӶڽϵ;ʶʵЧϵͳʶ.
ؼʣʶ𣬾ʶѧȣ䣬״̬פ
ͼţTP391.4
HMM STATE AND STATE DURATION DEPENDENT CONFIDENCE
MEASURES FOR UTTERANCE REJECTION
TIAN Bin, TIAN Hong-Xin, LIU Dan-Ting, and YI KeChu
(National Key Laboratory on Integrated Service Network, Xidian University, Xi'an 710071)
AbstractUtterance rejection is becoming increasingly important as speech recognition systems continuously migrate from the laboratory to actual applications. Proposed in this paper are state and state duration dependent acoustic confidence measures for acceptance/rejection of recognition hypothesis in speech recognition systems based on hidden Markov model (HMM). The state dependent confidence measure is computed for each frame of speech as the feature vectors output probability or posteriori state probability given the observation features. It is easy to be implemented by using one single global threshold and no extra training is needed. The state duration dependent one is based on the duration distribution probability and confidence interval theory. Although it is required that the state duration distribution be trained, the data can be easily obtained during the traditional HMM training. Experiment results show that the methods can reject incorrect candidates and OOV (out of vocabulary) words effectively, thus significantly increasing the recognition accuracy with low rejection rate.
Key wordsspeech recognition, utterance rejection, acoustic confidence measure, confidence interval, state duration
1
ʶ͹ؼȷϵͳУʶһǳҪ.ʶУѧʶʲ㹻ߣ1,2ݣѧƥõĴѡģʹܻõϸߵĸֵյõĽ[3]бҪþʶֶȥԴĺѡ4,5ݣͬʱʶЧؽͺѡĸ·.ڹؼȷϵͳУʶܹϵͳʣʹϵͳմʱ(OOV)ʹϵͳУʶʱҪ߿ɿʶǳҪ.ЩϵͳʶѾ൱ߣ95%ңʵûڲ࣬ģʶ۴ϻǴӼ϶һ.ˣԾʶʻȡʶʵ½ѾΪʶϵͳʵûһЧ;.
ʶһΪ࣬һýģ(filler)ͷؼģͣ67ǷʱڵĴʵо.ַҪԾʶѵרŵģͣһֱʶȻֵо89.ǰߵѵӣһֱģʽʶеľʶԭݹﲻǰߵˮƽ.
ľʶ㷨ںߣ϶ʶеƥϢõϺõĽ.״̬ʸƽ۲ʺ͸ʸ״̬ĺʾȽ趨ͳһľʶޣҲרŵģѵ.״̬פֲطǻפֲʺۣ趨һʶޣͬʱɸʶѡ״̬פŶ.
2״̬ʸصѧ׼
һHMMʶϵͳȷѡģƥɺƥ䣬Ҹ״̬ӦǺϵýϺ.ٶһHMM ={,A,B}S={s1,s2,sN}N״̬.Ϊʼ״ֲ̬ABĶɲμף10.O={O1,O2,,OT}Ϊһض۲ʸ.Viterbi룬ɵõOӦ״̬мO={Ob1,,Oe1,Ob2,,Oe2,,ObN,,OeN},biݺeiݷֱΪ״̬iӦʼֹ֡.HMMƥ÷ֲö÷ʱɵõѧŵ÷.
2.1״̬һ۲ֲ
(1)
÷Ա״̬פ֡ƽפ״̬÷ûפ̵״̬÷.
2.2״̬һֲ
(2)

(3)
һ״̬ڸ۲ʸµĺ.£HMMѧŵ÷ֶ״̬ƽ.־ʶ㷨ŵڲΪʶѵµģ.
3 ״̬פֲطѧ׼
״̬פģͳڸĽHMMģͶ״̬פȱ11,12,ȴʹHMMģ͵ѵʶ𶼱ø.õѵפֲʶб.ٶ״̬פdϸ˹ֲ(ǵϵͳô˷ֲ)Pi(d)=N(d,i,2i)٤ֲViterbi룬ɵõOӦ״̬еפ֡dj,j=1,2,,N.ɵó״̬פֲŵ÷
(4)
һ׼ȻOOVʶʶܽϺãЧACS1ACS21.֣ʶ෢ƽפر̵״̬.ʱ ,ܶȺֵܴܺ,С,Ĺ׾Եòȶƽפϳ״̬Ĺ׶ȶ̵СһЩ,Ȼʵ.ڸĽHMMģͶ״̬פĸģʵҲͬ.
1Ŷھʶʵ
ʶ׼ʶ(%)ʶ(%)ʶȷ(%)ʶ(%)
0.0%93.17*
2.5%95.5399.200.80
ACS15.0%97.7094.006.00
7.5%99.8098.801.20
2.5%95.5198.401.60
ACS25.0%97.7093.406.60
7.5%99.7098.801.20
2.5%95.3491.6010.05
ACS35.0%97.4092.008.00
7.5%99.5098.501.50
2.5%95.5399.200.80
ACS45.0%97.9093.403.60
7.5%99.8098.821.18
2.5%95.5699.600.40
ACS55.0%97.9097.402.60
7.5%99.8098.831.17
* 粻200ʱʻ㣬ʶΪ95.1% 
ֲ̬Լ95%ݷֲٽۦ-2ң+2ң.Ҹ䣬ɸֲ̬µֵӦŶ.ȷСλҲ500ֵʵ.ˣѧŵ÷.
(5)

(6)
kĴСƲ.
(7)

(8)
ʽ(7)ʾܸúѡԸúѡΪ߽Ŀ.
4ʵ
16 KHzݽʵ飬ʶΪ14άMelƵʵϵ(MFCC)֡Ϊ480(30 ms)֡160(10 ms)ʵһ49ʵķضʶϵͳн.ʶ22˹200200OOV(һЩʹʱ׻Ĵʺͼ).ļʶڲͬʶ½ǶԵһѡڽнܻоõ1Ľ.ڱ1
(9)
(10)
(11)
(12)
5ۺ
ķкܺõľʶܣܹϺõؾʶʱʶĽڽϵ;ʶʵ¿Чϵͳʶ.ACS3ЧACS1ACS2,Ϊĳ״̬ƽפܶʱ,ܶȺֵܴܺ,С,ôĹ׾Եòȶ.۵ACS4ACS5ȴ˷һȱ.״̬ʸصѧ׼Ҫרŵѵͽģ״̬פطµѵΪ״̬פģݿԴԭHMMѵеõѵȽϼ.׼򶼱Ƚþʶ.Ҫ˵ǣʵУӦĿĵĲͬҪʶʺ;ʶ֮.ͬʱע⵽ʶȷʺ;ʶƺ;ʶʲںֱӵĹϵͨԴڲͬ׼׻ʺOOVĵ÷Բͬ.ʾǽͬۺӦÿܵõõЧ.ӦһоԼڽͺѡĿʶмռ.ĵķHMMģͶ״̬פĸָĽģһοֵ.
*õѧӾϢʵһ.
߼飺У197010ʦʿоҪоΪʶ.
ģУ19683ʦҪоΪͨźŴ.
ͤУ19715˶ʿоҪоΪʶ.
׿˳У194310ڣʿʦҪоΪͨ.
ߵλӿƼѧۺҵصʵҡ710071
ο
1Fu S W K, Lee C H, Clubb O L. A survey on chinese speech recognition communication. Communication of COLIPS, 1996, 6(1): 117
2Makhoul J, Schwartz R. The voice of the computer.IEEE spectrum, 1997,34(12):3947
3Tian Bin, Tian Hongxin, Fu Qiang et al. A novel statistical language modeling method for continuous chinese speech recognition.In: IEEE Int'l Conf on Signal Processing, Beijing, 1998, vol I, 734737
4Renals S.Phone deactivation pruning in large vocabulary continuous speech recognition. IEEE Trans on Signal Processing Letters, 1996, 3(1):46
5Sukkar R A, Lee C H. Vocabulary independent discriminative utterance verification for nonkeyword rejection in subword based speech recognition. IEEE Trans on Speech and Audio Processing, 1996,4(6):420429
6Wilpon J G,Rabiner L R,Lee C H et al. Automatic recognition of keywords in unstrained speech using hidden markov models. IEEE Trans Acoustic, Speech, Signal Processing, 1990, 138(11): 18701878
7Hueng E F, Wong H C, Soong F K. A fast algorithm for large vocabulary keyword spotting application. IEEE Trans Speech  Audio Processing, 1994, SAP2(3): 449452
8Bourlard B, Hoore B D, Boite J M. Optimizing recognition and rejection performance in wordspotting systems. In: 1994 IEEE Int'l Conf on ASSP, Adelaide, Australia, 1994, vol I, 373376
9Rivlin Z. A confidence measure for acoustic likelihood scores. In: Proc of Eurospeech, 1995, Vol I, 523526
10 Rabiner L R. A tutorial on hidden markov models and selected applications in speech recognition. Proc of the IEEE, 1989, 77(2): 257286
11 Burshtein D. Robust parametric modeling of duration in hidden markov models. IEEE Trans on Speech  Audio Processing, 1996,4(3): 240242
12 սӢ½. ʶɷģ͵ĸĽ. ѧ199422(1): 915
(Zhan Puming, Wang Zuoying Lu Dajin. Improvement of hidden markov model for speech recognition, Acta Electronica Sinica(in Chinese), 1994, 22(1): 915)
ԭյڣ1998-12-07޸ĸյڣ1999-03-22.
