Acoustic-phonetic modeling in the SPICOS system

This paper deals with the acoustic-phonetic modeling developed in the SPICOS recognition system. The phoneme units are represented by variants of hidden Markov models and are refined step-by-step to improve the recognition performance. Two different approaches to modeling the emission probabilities are investigated, namely discrete models and continuous mixture densities. For the mixture density approach, a straightforward integration into the Viterbi scoring criterion is derived. In the experimental tests, the continuous mixture density approach was found to be superior to the discrete models, given the constraints of the Viterbi framework. The authors describe the details and refinements of the two approaches. >