Application of Discriminative Methods for Isolated Word Recognition
This paper describes a Hidden Markov Model based system for automatic recognition of isolated digits over telephone lines. For an LDA based linear feature transformation the classes to discriminate are choosen to be the HMM states. For MCE training this selection of classes is compared to the usage of the lexical words treated as classes. Experiments show that for MCE based reestimation of model parameters the latter choice is more appropriate, although in the case of Maximum Likelihood trainined parameters the correlation between Word Error rate and State Error rate is quite high.
KeywordsFeature Vector Linear Discriminant Analysis Word Error Rate Minimum Classification Error Linear Discriminant Analysis Class
Unable to display preview. Download preview PDF.
- BAHL, L.R. et al. (1993): Estimating Hidden Markov Models Parameters So As To Maximize Speech Recognition Accuracy. IEEE Transactions on SAAP, Voll, NO. 1, January 1993 Google Scholar
- CHO, W., JUANG, B.H., LEE, C.H. (1992): Segmental GPD Training of HMM Based Speech Recognizer. Proceedings ICASSP 92 Google Scholar
- HAEB-UMBACH, R., GELLER, D., NEY, H. (1993): Improvements in Connected Digit Recognition Using Linear Discriminant Analysis And Mixture Densities. Proceedings ICASSP 1993 Google Scholar
- HAUENSTEIN, A., MARSCHALL, E. (1995): Methods for Improved Speech Recognition over Telephone Lines. Proceedings ICASSP 95 Google Scholar
- KLISCH, R. (1996): The Voice Mail Digits and Their Performance on ICSI’s Hybrid HMM/ANN System. technical report, International Computer Science Institute Google Scholar