A probabilistic state machine for speech recognition

This paper proposes a recognition structure designed for handling continuous speech in a natural and computationally efficient way. without the need for a higher Ievel algorithm (like. e.g .• Ievel building). This structure is based on a probabilistic state machine (PSM). but unlike Hidden Markov Models. the transition probabilities at each time frame depend on the observation made on the input speech signal. in that frame. Some of the states of the PSM are associated to the various words to be recognized. such that a high probability in one of those states at a given time is interpreted asa high probability that the corresponding word to that state has been found. at that time. in the input signal. This model is highly efficient, requiring only one vector-matrix multiplication per input Observation. The theoretical formulations of the recognition and training algorithms are presented. together with some very preliminary experimental results.