Minimum classification error rate method using genetic algorithms

Hidden Markov models (HMM) is one the most common statistical matching methods used for speech recognition, especially for continuous speech utterances. One major problem in HMM is that the training process that aims to generate a set of HMM models (recognizer) for matching the speech source usually is based on a set of limited training data. The minimum classification error (MCE) training approach (Juang et al., 1997) is regarded as a discriminative method that is proven to be superior to other traditional probability distribution estimation approaches, such as the maximum likelihood (ML) approach. The main drawback in the MCE is to formulate the error rate estimate function as a smooth loss function for applying a gradient search technique that subsequently leads to a local optimal solution. In this paper, a genetic algorithm based approach (GA-MCE) for the MCE is proposed to solve these problems. In our experiments, the results demonstrated that the GA-MCE is superior to the original MCE method. It can also significantly increase the performance of voice input systems.