Speech recognition based on k-means clustering and neural network ensembles

Aiming at the disadvantages of the single BP neural network in speech recognition, a method of speech recognition based on k-means clustering and neural network ensembles is presented in this paper. At first, a number of individual neural networks are trained, and then the k-means clustering algorithm is used to select a part of the trained individuals' weights and thresholds for improving diversity. After that, the individuals of the nearest clustering center are selected to make up the membership's initial weights and thresholds of the ensemble learning. The method not only overcomes the shortcomings that single BP neural network model is easy to local convergence and is lack of stability, but also solves the problems that the traditional adaboost method in training time is too long and the diversity of individual network is not obvious. The final experiment results prove the effectiveness of this method when applied to speakers of independent speech recognition.