A new mixture of experts framework for acoustic modelling using SOM clustering and radial basis funtions

In machine learning applications, Supervised Mixture models and Mixture of experts play vital role in performing pattern classification or phoneme classification eg. acoustic modelling of speech recognition. In this paper, we introduce a new mixture of experts' classification kernel by embedding self organized map (SOM) clustering with mixture of radial basis function (RBF) networks. The model's efficacy is demonstrated in solving a multi-class TIMIT speech recognition problem where the kernel is used to learn the multidimensional cepstral feature vectors to estimate their posterior class probabilities. The tests results have shown that this model provides a better alternative to the state of the art models achieving a significant improvement in error performance, reduction in complexity and gain in training time.