A knowledge based approach for automatic labeling of a large speech database

Abstract

The authors describe a novel automatic segmentation system based on perceptual cues, spectral dynamics, and various sources of knowledge: heuristic, phonetic, and suprasegmental. Because no reference units are used, the method can be directly applied to speech recognition. The passage from one application to another is facilitated by the modularity and openness given by the backboard model. This method has several advantages over other segmentation methods presented in the literature: it is based on an open and modular architecture; the smooth spectrum yielded by the perceptually based linear prediction analysis limits spurious segments; the algorithm which determines the transitions uses no threshold and thus is speaker independent; the system does not require reference units; and the introduction of phonetic knowledge deals mostly with the difficult cases.<>