In 5. References Mutual Information Estimation and the Speech Re

the following optimization of the discriminative functions the extended BW algorithm (12) was applied to re-estimate the mixtures of the phoneme models. All phoneme alternatives within the automatically derived phoneme seg-mentation were used in the calculation of the discriminance measure. According to (6), the mixtures of the correct and of all competing models were updated to minimize the objective function. Best results were obtained in MCE optimization using the update equation (15) with = 1:0, which results in a stable learning and a phoneme recognition rate of 64.8 % on the test set. For = 0:3 objective function and error rate were decreasing monotonously, but recognition performance within the 5 iterations was slightly worse (62.9 %). MMI training, using the extended BW algorithm, was stable for a smoothing parameter = 0:3, in which the objective function and the recognition rate were increasing monotonously. The MMI objective is less robust than the MCE function and therefore requires smaller`step-sizes'. 62.0 % phoneme recognition rate were achieved by the MMI optimization, which is an improvement of 2.5 points compared to the ML baseline system. In these experiments MCE training was more stable than MMI optimization and resulted in higher phoneme recognition results. Furthermore an alternative calculation of the discrimi-nance measure based on the best phoneme sequence hypothesis for the utterance was applied. Now the correct description versus the best hypothesis of the utterance, which was derived by a looped phoneme model without a lexicon or language model, was used to compute the objective function. Since both descriptions diier only in some parts of the sentence, just these diierent segments were actually used in discriminative training. Therefore only small parts of the training data were used in the discriminative parameter re-estimation process. Only MCE optimization was examined for this technique, which resulted in minor improvements of 1.7 points to 61.2 % phoneme recognition rate within the same number of iterations. Using`N-best' alternative hypothesis would improve this training scheme by generating more competing hypotheses for the discriminative training. 4. CONCLUSIONS In this paper a discussion of the MCE and the MMI objective was presented. For the HMM parameter update an extended BW reestimation formula was suggested, which can be used for both discriminative methods. It was applied in speaker independent phoneme recognition experiments and improved the recognition rate about 5.3 points from 59.5 % to 64.8 % for the MCE function. In our experiments MCE …