An improved K-means clustering algorithm and application to combined multi-codebook/MLP neural network speech recognition

Unsupervised learning algorithms play a central part in models of neural computation. K-means clustering algorithms, a type of unsupervised learning algorithms, have been used in many application areas. We propose an improved K-means algorithm for optimal partition which can achieve better variation equalization than standard binary splitting algorithms. The proposed clustering algorithm was applied to combined multi-codebook/MLP neural network speech recognition system to train the LPC based codebooks. It achieved smaller variation of the variances of clusters than that from the standard binary splitting algorithm.