Speech recognition using dynamic transformation of phoneme templates depending on acoustic/phonetic environments

A description is given of a phoneme-based speech recognition method using a linear model to represent the contextual variations in acoustic features caused by the phonemic context. A feature vector in a phonetic segment is decomposed into a context-independent vector, a context-dependent vector, and a residual vector. The context-independent vector is calculated by weighting a coefficient matrix to a context vector obtained from acoustic and/or phonetic data dynamically. Algebraic formulas for the maximum-likelihood estimations of the parameters in the model are derived by statistical modeling of the residual vector. For example, the proposed model achieved 97.9% word accuracy while the whole-word template model obtained 95.1% word accuracy