Method and system for speech recognition by means of hidden Markov models with continuous output probabilities