A comparison of modified k-means (MKM) and NN based real time adaptive clustering algorithms for articulatory space codebook formation

The paper proposes the use of a neural network based real time adaptive clustering algorithm for the formation of a codebook of limited set of acoustical representation of finite set of vocal tract shapes from an articulatory space. A modified k-means algorithm (MKM) used for clustering nearly 10000 vocal tract shapes into 1000 cluster centers to form a codebook of articulatory shapes is computationally intensive for the application. An investigative study on the use of the NN based algorithm over the MKM algorithm at the peripheral level, for an application on computer aided pronunciation education, suggests the former for less intensive computation, with the possibility of improving the performance of the system by implementing the algorithm using a dedicated neural computer. Preliminary results of this study are reported.