(A) study on the clustering method in speaker independent isolated word recognition

Abstract

In this thesis, we propose a hybrid clustering algorithm called Modified Mutual Nearest Neighborhood(MMNN) clustering, which is aimed at superior performance to other clustering method for speaker independent isolated word recognition. This algorithm, compared to other clustering methods, Unsupervised Without Averaging (UWA) and Chainmap, shows better performance by generating more reliable reference patterns. A new cluster center determination method was devised. Its performance is superior to minimax. And the cluster center determined by this method is one of the real training pattern which saves many computation time.