`Quantized´ hidden Markov models for efficient recognition of cerebral palsy speech

A procedure for evaluating the likelihood of a hidden Markov model using only O(N/K) floating-point operations per observation is developed where N is the number of states in the model and K is a large integer. Experimental tests show that effective recognition of cerebral palsy speech requires highly connected models so that O(3N) to O(N²) operations are necessary using conventional algorithms. The reduction in computational complexity is required for near-real-time recognition algorithms to be feasible on ordinary personal computers