In this paper we describe the use of demisyllables as reference units in a speaker-dependent connected speech recognition system using template matching techniques. Without explicitly segmenting the input utterance, the algorithm finds out the sequence of demisyllables which best matches the input string. The use of syntactic constraints reduces the set of possible outputs to those gramatically correct. The system works on the first 1000 Spanish numbers. Preliminary experiments on demisyllable segmentation and digit recognition are described.

Peer Reviewed