Abstract

This thesis is part of the 'speech recognition' project. The Technical University Delft started just recently in the area of speech recognition. The project is in the phase to gather knowledge about speech recognition. For that matter the possibilities in speech recognition with the use of a Kohonen neural network on a nCUBE2 parallel computer are studied in this thesis. One of the applications presented by the designer of the Kohonen network is in the area of speech recognition, it concerns the recognition of Finnish phonemes. This report describes a speaker independent recognizer that is build using selforganizing techniques described by T. Kohonen. The implementation is executed on a parallel computer. The reason for this is straightforward, the final goal of this project is to build a real-time continuos automatic speech recognition system. Because such systems consist of time costly algorithms, parallel methods are examined to overcome the problems of time critical algorithms. A new approach for training a speech recognition system with the use of self organization is presented. This metod give promising results, and the final goal of building an automated continuos speech recognition system becomes closer.

To study the results of the implemented phoneme recognizer, the total set of recognizable phonemes were divided into vowels, fricatives, nasals and stops. The recognition results were statistically analyzed. The recognition performance showed good, but no excellence results. When the results were examined closer, clusters were found in the phonemes that were directly related phoneme classes. These clusters consist of phonemes that are mostly confused with each other. More research is however necessary, to find more conclusive evidence for this.