[1]王雨轩,倪训博,姜 峰.手语识别中基于HMM的区分性训练方法[J].智能系统学报,2007,2(1):80-84.
WANG Yu-xuan,NI Xun-bo,JIANG Feng.Discriminative training methods of HMM for sign language recognition[J].CAAI Transactions on Intelligent Systems,2007,2(1):80-84.
点击复制
WANG Yu-xuan,NI Xun-bo,JIANG Feng.Discriminative training methods of HMM for sign language recognition[J].CAAI Transactions on Intelligent Systems,2007,2(1):80-84.
手语识别中基于HMM的区分性训练方法
《智能系统学报》[ISSN 1673-4785/CN 23-1538/TP] 卷:
2
期数:
2007年第1期
页码:
80-84
栏目:
学术论文—机器感知与模式识别
出版日期:
2007-02-25
- Title:
- Discriminative training methods of HMM for sign language recognition
- 文章编号:
- 1673-4785(2007)01-0080-05
- 文献标志码:
- A
- 摘要:
- 传统的隐马尔科夫模型（HMM）的训练方法基于统计概率的最大似然准则（MLE），在训练样本数目足够大的情况下，这种方法在理论上可以得到最优的结果.在手语识别研究中，采集足够大的训练样本十分困难.区分性训练可以很好地弥补由于训练样本的缺乏以及手语模型之间的近似而造成的识别系统的缺陷.最大交互信息准则（MMIE）作为区分性训练准则的一种已经被广泛的应用于语音识别领域.文中通过合理的构建手语识别中的竞争模型和易混集，提出了MMIE准则的改进形式，并将其应用于特定人与非特定人手语识别.实验证明，使用改进的MMIE准则对识别系统性能有很大的提高.
- Abstract:
- The traditional method of training HMM (Hidden Markov Models) is based on MLE (maximum likelihood estimation). When training samples are sufficient en ough, the method can principally gain the optimal result. However, it is too dif ficult to get such large data sets practically, especially in sign language reco gnition. Discriminative training method can improve the error rate of MLE, which is caused by insufficient training data and similarities among sign language mo dels. Maximum mutual information estimation as one of discriminative training me thods has been widely applied in speech recognition. By taking competition model s into account and setting up mixture sets appropriately, MMIE method was improv ed and applied both in signerdependent and signerindependent sign language rec ognition. A great number of experiments had been taken, showing that this method greatly promoted the ability of the traditional MLE system.
- 参考文献/References:
-
［1］ WANG Chunli, GAO Wen. Resampling for Chinese sign language recognition by genetic algorithm［A］.GW2005［C］.［s.l.］，2005.
［2］DENG J W, TSUI H T. A twostep approach based on PaHMM for the recognition o f ASL［A］. Proceedings of The Fifth Asian Conference on Computer Vision［C］. M elbourne, Australia, 2002.
［3］BAHLL R, BROWN P F, SOUZA P V, MERCER R L. Maximum mutual information esti ma tion of hidden Markov model parameters for speech recognition［A］. Proc . 1986 I nt. Conf. on Acoustics,Speech and Signal Processing ［C］.Tokyo,Japan1986.
?［4］NORMANDIN Y. An improved MMIE training algorithm for speaker independent［ A］. Proc. ICASSP’91［C］. Toronto,1991.
［5］SCHLUTER R,MACHEREY W,RULLER B, NEY H. Comparison of discriminat ive training criteria and optimization methods for speech recognition［J］. Spee ch Communication, 2001(34):287-310.
［6］ZHENG J，BUTZBERER J，FRANCO H. Scandinavia impro ved maximum mutual information es timation training of continuous density HMMs［J］.Andreas Stolcke Speech Technol ogy and Research Laboratory，2001,15(2):25-30.
［7］WOODLAND P C, POVEY D. Large scale discriminative training for speec h reco gnition［J］. In Proc. ITRW ASR［C］. ISCA, 2000.
［8］BAHL L R，PADMANABHAN M,NAHAMOO D,GOPALAKRISHNAN P S. An nbest candidatesbased discriminative training for speech recognition Applications［J］.IEEE Transactions on Speech and Audio Proce ssing, 1994,2(1):206-216.
［9］CHOW Y L.Maximum mutual information estimation of HMM paramet ers for continuous speech recognition using the NBest algorithm［A］. Proc. IC ASSP ’90［C］. Albuquerque，1990.
备注/Memo
收稿日期：2006-04-29
.作者简介：
王雨轩，男，1980年生，哈尔滨工业大学硕士研究生，主要研究方向为模式识别、机器学习．E-mail:yxwang@vilab.hit.edu.cn.
?倪训博，男,1978年生，哈尔滨工业大学博士研究生，主要研究方向为模式识别、机器学习． E-mail: nixunbo@hit.edu.cn
?姜 峰，男，1978年生，哈尔滨工业大学讲师，主要研究方向为模式识别、机器学习、图像处理、人机交互等． E-mail: fjiang@hit.edu.cn
更新日期/Last Update:
2009-05-05