This paper describes the field trial of a speech system with an isolated word recognizer (HMM based) that offers a 'News Service' to the general public over the Spanish telephone network. The results of this trial show that real recognition rates are substantially lower that laboratory recognition rates, due mainly to human factor aspects, e. g. users do always not follow system instructions, speaking fluently instead of uttering isolated words when the system prompts them. To cope with this practical situation, a connected word recognition algorithm has been implemented, and some parametric models have been computed to perform keyword spotting. The algorithm and the way this models have been trained will be presented, and the results of the new system will be exposed. Keywords: Hidden Markov Models, Word Spotting, Telephone line, Speech Machine
Bibliographic reference. Poza, M. J. / Torre, C. de la / Tapias, D. / Villarrubia, L. (1991): "An approach to automatic recognition of keywords in unconstrained speech using parametric models", In EUROSPEECH-1991, 471-474.