A self-organizing clustering technique for vector quantization in speech recognition


For the sake of data reduction in automatic speech recognition often vector quantization based on a previously generated code book is performed. ln the approach described here the necessary code book is set up by means of a selforganizing dustering technique. lt takes the shape of a two-dimensional array of feature vectors. Phonetically similar vectors are also arranged in geometrical vicinity. The definition of a new distance measure suitable for this so-called phonotopic map is introduced. The procedure has been implemented for an isolated-word recognition system for I arge vocabularies (1 ,000 words). From a small number of phonetically balanced training utterances (17 words) a map of size 10x10 is built. A recognition rate of more than 98 per cent is achieved with single training of the lexicon when the phonotopic map is used as code book in combination with the proposed distance measure.

2 Figures and Tables

Cite this paper

@inproceedings{Aktas1987ASC, title={A self-organizing clustering technique for vector quantization in speech recognition}, author={A. Aktas and L. Gla\sser and Bernhard R. K{\"a}mmerer and Wolfgang A. K{\"{u}pper}, booktitle={ECST}, year={1987} }