Research On Discriminative Training For Speech Recognition

As a high-tech applications technology research hotspot, speech recognition technology has made great strides in recent years. The traditional acoustic model training is based on the maximum likelihood criterion. It just uses the training samples within each class to train class model parameters, so there exists some limits. The discriminative training will consider the discriminative information to adjust the classification surface between different models to gain better model performance. Therefore, the research on discriminative training is a very meaningful topic.Because of the traditional maximum likelihood estimates of automatic speech recognition systems’limits, the discriminative training was born. The research is mainly about the minimum phone error criterion, including the objective function of maximum algorithm (EM algorithm) and the model parameter update algorithm (Baum-Welch algorithm). It also introduces other associated techniques, such as smoothing technique, Gaussian mixture components. We conducted a series of experiments which proved the discriminative training has obvious advantages.The further optimization of discriminative training includes two aspects. One is to definite the objective function, which aims to propose new training guidelines to make sure the optimization indicators. The other one is to optimize the training algorithm which is about the process how to get the optimal objective function’s model parameters. In the paper, the mainly research is as following.1. Aimed to optimize the objective function, the paper proposed a new minimum phone error objective function (Time measure MPE). It associated with the F1function which is one kind of confidence function to update the traditional objective function’s risk function. It can remedy the traditional objective function’s limits of ignorance of deletion error.. There is an experiment which proved the feasibility of the algorithm.2. Considering to optimize the training model, we can use discriminative training criteria to guide feature extraction and further update the acoustic mode. Its purpose is to adjust the feature to gain the optimal objective function. It gains the new feature through feature conversion and then uses discriminative training to optimize the model. Related research includes high-dimensional feature vectors, the transformation matrix and so on. Through the implement of the algorithm, it proved that it can improve the performance of the model.