The `ARMADA´ continuous speech recognition system

Hidden Markov Models (HMMs) provide a formal framework for statistical modelling of time-varying patterns such as speech patterns. Implicit in the use of HMMs is the assumption that these patterns consist of a sequence of quasi-stationary segments, and that the sequence of feature vectors in each segment can be modelled as the output of a stationary stochastic process. This assumption is clearly inaccurate in the context of speech patterns, which vary continuously according to the movements in the speaker vocal apparatus. However the limitations of HMMs are offset by the availability of proven mathematical techniques for automatically estimating the parameters of a set of HMMs from example speech data, and for classifying an unknown speech pattern with respect to a set of HMMs. The author looks at the ARMADA speech recognition system which consists of approximately 1500 HMMs