Abstract.
FAPES system is based on a specialized Fuzzy ARTMAP network trained to estimate the observation probabilities NN-HMM speech recognition systems require. This Fuzzy ARTMAP classifier transfers after ARTa resonance the choice function of all eligible nodes to a SLP defuzzifier, that was trained to map fuzzy scores to a-posteriori probabilities of visiting HMM states. Significant computing time reduction results from estimating observation probabilities with such local error propagation NN, instead of well-known Multilayer Perceptron. Furthermore, Fuzzy ARTMAP estimator determines inherent discrimination among output a-posteriori probabilities, discrimination that can only de added into HMM training by means of complex MMIE algorithm. Input feature space partitioning is also considered to emphasize the weight CEPS coefficients and short-term energy subset has on overall state probability. An iterative training procedure has been used to instruct Fuzzy ARTMAP and SLP components, no training being required for the HMM part as transition probabilities are negligible. Following new ideas were introduced to proposed training procedure: the number of ARTa clusters is asymptotically growing to some imposed limit. SLP defuzzifier training algorithm is inferred from the simplified block of IF-THEN fuzzy rule controller paradigm.no global lateral inhibition leading to winner selection in F2 layer of ARTa subnet occurs. It is thus possible that all eligible nodes send their choice function to SLP defuzzifier.