Semi-automatic phonemic labelling of speech data using a self-organising neural network

In the perspective of assessing existing and new speech input and output devices, the development of methods for computerised semi-automatic labelling of speech data is becoming increasingly important. This paper describes preliminary work to achieving this goal by utilising Neural Network technique to perform phonemic classification. The speech data used for Neural Network learning are taken from the SAM-EUROM.O database, which has been manually labelled by expert phoneticians as a reference. Separate data from 1 male speaker are used to test the performance of the semiautomatic phonemic labelling system. The preliminary results show an average classification rate of 65 % an the vocalic and 72% an the consonantal Danish archiphonemes.