Towards better understanding of the model implied by the use of dynamic features in HMMs

We examine a widely-used kind of Hidden Markov Model (HMM), in which “dynamic features” are included along with the direct measurements. We conclude that the generative model implied by the use of dynamic features is quite different from the conventional view and that such models are capable of providing a surprising amount of the sort of dynamics that we thought were necessary to describe the important properties of speech patterns. We suggest that one reason why several attempts to replace HMMs with models with explicit dynamics have failed is that the dynamics already implicit in standard HMMs are roughly equivalent.