This is the 1st year funding of a 3 year continuing award. This research is to develop a new, model-based, spectral estimation algorithm for recognition of noisy speech. The new algorithm represents a significant improvement over previous estimation algorithms because it incorporates more information about speech spectral distribution. The proposed work improves the estimation by incorporating information about its dynamic properties and its quasi-periodic nature. To capture the dynamic properties the research will study several types of hidden Markov models. An estimation of the position of the harmonics will be made to address the periodicity properties. From these will be made an estimate of the spectral energy at any given frequency dependent on its proximity to the nearest harmonic. Incorporation of dynamics and periodicity should improve speech recognition performance in noisy environments.
