P.C. Chang, B.H. Juang, "Discriminative Training of Dynamic Programming Based Speech Recognizers," IEEE Trans. Speech and Audio Process., pp.135143, volume 1, number 2, April 1993.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....in practical speech recognition problems, neither of the above assumptions hold. Thus, there is an (at least theoretical) advantage to direct optimization of the MCE rather than the ML criterion. The shortcomings of ML have been recognized by several researchers (e.g. 1] 2] 6] 18] [7]) who provided strong 1063 6676 01 10.00 2001 IEEE reasons for discarding it in favor of discriminative design methods [31, Sec 5.6] that jointly optimize all the HMMs in a classifier. One promising discriminative design method, Generalized Probabilistic Descent (GPD) was proposed and extended ....

....a smooth and differentiable function. Once the cost is smoothed, gradient methods may be used to optimize the classifier s parameter set and find a local minimum on the (smoothed) surface. Discriminative methods have been applied to the design of speech recognizers based on both template matching [7], 24] and HMMs [9] 18] In the context of HMMs, it was shown in [9] and [18] that GPD provides a significant improvement in recognition accuracy over ML design. Our starting point, however, is with the observation that while cost surface smoothing allows the use of gradient descent ....

[Article contains additional citation context not shown here]

P.-C. Chang and B.-H. Juang, "Discriminative training of dynamic programming based speech recognizers," IEEE Trans. Speech Audio Processing, vol. 1, pp. 135--143, Apr. 1993.

.... This application of MCE based adaptation to speech recognition using DTW was evaluated on the Bell Labs E set task, and on phoneme recognition tasks for the ATR 5240 word databases, with good results [37 38] A very similar application of MCE to DTW based speech recognition was described in [39 40], which yielded similar results for the E set task. 5.7.2 Prototype based Minimum Error Classifier (PBMEC) A different approach to DP based speech recognition using speech pattern prototypes is the Prototype based Minimum Error Classifier (PBMEC) ar chitecture proposed in [41 42] The difference ....

P.-C. Chang and B.-H. Juang. Discriminative Training of Dynamic Programming Based Speech Recognizers. IEEE Transactions on Speech and Audio Processing, 1(2):135-143, 1993.

....the perspective of increasing classifier 59 robustness. In parallel with this work, originally reported in [McDermott Katagiri, 1991b] the MCE GPD framework has been applied to speech recognition problems in several different ways, including the methods described in [Komori Katagiri, 1991, Chang Juang, 1993a, Rainton Sagayama, 1992b, Chou et al. 1992] The differences between these classifiers and PBMEC will be described in this chapter. 4.2 Prototype Based Minimum Error Classifier (PBMEC) 4.2.1 Arbitrary Unit Level Training In Chapter 3, the Shift Tolerant LVQ (ST LVQ) architecture was applied ....

....single segments of speech. For these reasons, the new term, Prototype Based Minimum Error Classifier (PBMEC) was introduced to refer to this idea. DTW classifiers The structure of the classifier described here is quite different from that of MCE GPD trained DTW classifiers [Chang Juang, 1991, Chang Juang, 1993a] and [Komori Katagiri, 1992a, 1992c] These approaches use full templates of words or phonemes, whereas the PBMEC system described here uses a small number of sub word or sub phoneme states to represent a given speech category. These states function as abstractions of adjacent frames in a ....

[Article contains additional citation context not shown here]

Chang, P.-C. and Juang, B.-H. (1993). Discriminative Training of Dynamic Programming Based Speech Recognizers. IEEE Transactions on Speech and Audio Processing, Vol. 1, No. 2, pp. 135-143.

....is at the level of single segments of speech. For these reasons, the new term, Prototype Based Minimum Error Classifier (PBMEC) was introduced to refer to this idea. DTW classifiers The structure of the classifier described here is quite different from that of MCE GPD trained DTW classifiers [Chang Juang, 1991, Chang Juang, 1993a] and [Komori Katagiri, 1992a, 1992c] These approaches use full templates of words or phonemes, whereas the PBMEC system described here uses a small number of sub word or sub phoneme states to represent a given speech category. These states function as abstractions of ....

....when correct, as determined by the value of the threshold Q1 in the above loss function, are described. 4.3. 2 Overview of Experiments The experiments described here were designed to test the capability of PBMEC on databases that have been used by others, namely the Bell Labs E set task [Chang Juang, 1991] and the ATR 5240 isolated word recognition task [McDermott Katagiri, 1991b] On both the E set task and the 5240 isolated word recognition task, the effect of varying the loss function parameter Q1 was examined. For both speaker dependent and multi speaker experiments, it was found that this ....

[Article contains additional citation context not shown here]

Chang, P.-C. and Juang, B.-H. (1991). Discriminative Training of Dynamic Programming Based Speech Recognizers. Proceedings of the IEEE, ICASSP-91, pp. 549552.

....whose objective is minimum classification error (MCE) MCE techniques smooth the classification error cost function and jointly optimize all HMM parameters of the classifier via gradient descent. Of particular importance within the MCE family is the generalized probabilistic descent (GPD) [1,2,3,5]. Although MCE targets the true design cost and thereby offers significant performance gains over ML, it suffers from a significant drawback. The MCE cost surface is riddled with shallow local minima that easily trap local descent methods, and may substantially compromise performance. The above ....

Chang P.-C. and Juang B.-H. (1993), Discriminative training of dynamic programming based speech recognizers. IEEE Trans. On Speech and Audio Processing, Vol. 1, No 2, April 1993, pp 135-143.

No context found.

P.C. Chang, B.H. Juang, "Discriminative Training of Dynamic Programming Based Speech Recognizers," IEEE Trans. Speech and Audio Process., pp.135143, volume 1, number 2, April 1993.

No context found.

P. C. Chang and B. H. Juang, "Discriminative training of dynamic programming based speech recognizers," IEEE Trans. Speech Audio Processing, vol. 1, no. 2, pp. 135--143, 1993.

No context found.

P. C. Chang and B. H. Juang, "Discriminative training of dynamic programming based speech recognizers," IEEE Trans. Speech Audio Processing, vol. 1, pp. 135--143, 1993.

No context found.

Chang, P.-C. and Juang, B.-H. (1993). Discriminative Training of Dynamic Programming Based Speech Recognizers. IEEE Transactions on Speech and Audio Processing, Vol. 1, No. 2, April 1993. pp. 135-143.

No context found.

Chang, P.-C. and Juang, B.-H. (1991). Discriminative Training of Dynamic Programming Based Speech Recognizers. Proceedings of the IEEE ICASSP-91, pp. 549-552.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC