Humans can recognize different types of written languages by their grammars and vocabularies. However, computers see everything as numbers. We present a computational algorithm for machine classification of written languages using the method of vector quantization. For a language document, each word is converted to a sequence of numbers and forms as a vector of numerical values according to its characters. This collection of vectors is then represented by a codebook that contains a number of template vectors for classification. The proposed method is more effective for machine learning than the n-gram based method, which has been widely used for written language identification. Experimental results of classifying a set of five closely roman-typed scripts show the promising application of the proposed method.