In the framework of COST 232 an isolated word recognizer (called RECO) based on discrete-density HMMs (hidden Markov models) was developed. A series of optimizations resulted in a speaker-independent word recognition rate of 99 per cent for the German digits spoken over dialed-up telephone lines (see [Hut94]).
Further investigations have been made using hybrid HMM/ANN approaches, as can be seen in [HP94], [Hut95], and [Hut96].
Supported by: Bundesamt für Bildung und Wissenschaft