Phoneme Recognition for Speaker-Independent ConnectedWord Recognition

In this paper, a method of phoneme recognition is described for speaker-in-dependent connected word recognition. The phoneme recogmtIOn is carried out by two processes, namely, segmentation and phonemic labeling. In the segmentation, insensitivity is desirable to speaking rate or structural difference of speech production organs. To achieve such insensitivity, we propose an optimal segmentation technique by the variance-minimization which globally divides the input speech pattern into segments by minimizing the sum of the vari-ance of each segment, in stead of local and sequential division. In the phoneme labeling, segment labels are determined on the basis of the frame labels which are determined by template matching between phonemic refe-rence patterns and the frames. It is devised that the phonemic reference patterns to be matched with the frames are reduced using the inter-segment and intra-segment information. 1