Optimal spoken dialog control in hands-free medical information systems

In the paper a method of optimal selection of utterances used as command entry-words for voice controlled application is presented. Voice controlled programs seem to be particularly useful in the area of medical informatics, where a physician interacts with a program by voice while operating the medical device or being involved in examinations requiring manual activities. The proposed method selects command words from sets of proposals defined for each command so as to minimize the overall probability of incorrect command recognition. First the entry-word dissimilarity matrix is calculated. The word dissimilarities are evaluated using HMM models consisting of appropriately trained acoustic models of the phonemes constituting words. The trained HMM is used as the sample utterance generator for the word. The artificially created utterance samples are then recognized by speech recognizers created for pairs of words. The estimation of correct recognition probability is used as the word dissimilarity measure. The word dissimilarities are then used to determine the average assessment of words selections that can be used as commands. Selection is created by choosing single word from sets of candidates defined for each command. Finally, suboptimal selection is found by using genetic algorithm. Experiments carried out prove that suboptimal selection of command entry-words can observably increase the accuracy of spoken commands recognition in many cases.