Maximum-likelihood updates of HMM duration parameters for discriminative continuous speech recognition

Previous studies showed that a signi cantly enhanced recognition performance can be achieved by incorporating information about HMM duration along with the cepstral parameters. The reestimation formula for the duration parameters have been derived in the past using xed segmentation during K-means training and the duration statistics are always xed throughout the additional minimum string error (MSE) training process. In this study, we update the duration parameters along with other model parameters during discriminative training iterations. The convergence property of the training property based on the MSE approach is investigated, and experimental results on wireline connected digit recognition task demonstrated a 6% word error rate reduction by using the newly trained duration model parameters as compared to xed duartion parameters during MSE training.