Phone recognition in continuous speech (Dutch)

The use of hidden Markov models is a powerful tool in building automatic speech recognition systems for continuous speech. Instead of just using frame‐by‐frame information, dynamic information is also introduced by taking the time derivative of static parameter vectors. The static and dynamic parameters are processed in separate codebooks and these observations are then integrated at a probabilistic level [Gupta et al., Proc. ICASSP, 697–700 (1987)]. The single‐user, connected‐word recognition system is developed for a subvocabulary of the Dutch language (appropriate for banking transactions). The phone models (used as subword units) are made more robust by using the co‐occurrence smoothing algorithm [K. F. Lee and H. W. Hon, IEEE Trans. Acoust. Speech Signal Process. ASSP‐37(11), 1641–1648 (1989)], which enables accurate recognition, even with limited training data. Results will be presented at the meeting.