GPD-Based State Modification by Weighted Linear Loss Function

Abstract

This paper proposes an effective loss function to assign the HMM state weight based on the MCE method. This is to remedy the performance limitation inherent in the traditional maximum likelihood method, which adjusts parameters to maximize the likelihood of training HMM. If minimum classification error method is used to minimize the error rate for training data, then the local optimum point can be achieved and the recognition performance can be achieved. However, if the amount of data used in the MCE training is too small, there can be a risk of overfitting to the training data. In this paper, we propose state modification by weighted linear loss function to overcome overfitting to training data. Representative experiments confirm this postulation and show the improvement in error rate when applied.

These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.