University of Washington - Department of Statistics
A two-dimensional extension of Hidden Markov Models (HMM) is introduced, aiming at improving the modeling of speech signal spectrograms. The extended model:
-focuses on the conditional joint distribution of state durations given the length of utterances, rather than on state transition probabilities;
-extends the dependency of observation densities to current, as well as neighboring states; and
-introduces a local averaging procedure to smooth the outcome associated to transitions from successive states.
A set of iterative algorithms, based on segmental K-means and Iterative Conditional Modes, for the implementation of the extended model, is also presented. In applications to the recognition of segmented digits spoken over the telephone, the extended model, achieved about 23% reduction in the recognition error rate, when compared to the performance of HMMs.
This is joint work with Jiayu Li, Department of Statistics, University of Chicago.