A New Training Strategy for the Learning Subspace Method of Classification, Applied to Simultaneous Phonemic Segmentation and Labeling of Continuous Speech

摘要

A variant of the original learning subspace method, applied to simultaneous phonemic segmentation and labeling of continuous speech, is introduced in this paper. The learning ability in the basic method is accomplished by decision-controlled rotations of the subspaces into proper directions during the iterative training period, whereby in the steady-state regions of the phonemes, the projections of the prototype vectors are required to be largest onto the subspace corresponding to the correct phonemic subspace. A new feature of training strategy introduced here is to include transitional information of speech into the learning process which improves the segmentation accuracy.