Second International Conference on Spoken Language Processing (ICSLP'92)

Banff, Alberta, Canada
October 13-16, 1992

Single Word Detection System with a Neural Classifier for Recognizing Speech at Variable Levels of Background Noise

D. Barschdorff, U. Gartner

Electrical Measurement, University of Paderborn, Paderborn, Germany

This paper proposes a speech recognition system based on a neural network. The network with good generalising capabilities is used for speaker independent single word classification. The system is able to follow changes in the pronunciation by the same speaker and it can also be adapted to previously unlearnt speakers by a special retraining mode. Speaker independent recognition experiments have been performed using a vocabulary of 21 German words by male and female speakers. Using the proposed, retraining method, a high recognition accuracy of better than 90% is obtained after a few spoken words by an untrained speaker.

Full Paper

Bibliographic reference.  Barschdorff, D. / Gartner, U. (1992): "Single word detection system with a neural classifier for recognizing speech at variable levels of background noise", In ICSLP-1992, 1347-1350.