First International Conference on Spoken Language Processing (ICSLP 90)
Clustering is one of the most prevalent methods to construct multi-templates from training data. However, most of clustering algorithms proposed in the literature aim at minimizing not the recognition error, but the square error distortion measure. This paper describes a new clustering algorithm which optimally classifies training data into clusters in such a way that it minimizes a recognition error function. Optimization is accomplished by the simulated annealing technique. The new algorithm is compared with the LBG clustering algorithm and the LVQ2 algorithm in vowel template learning experiments, and confirmed to yield the best results. This paper also investigates a possibility of using an iterative improvement method as an optimization technique in the new algorithm.
Bibliographic reference. Ando, Akio / Ozeki, Kazuhiko (1990): "Clustering algorithms to minimize recognition error function and their applications to the vowel template learninig", In ICSLP-1990, 217-220.