A new segment-based Dynamic Programming (DP)/Multi-Layer Perceptron (MLP) hybrid for speaker-independent continuous speech recognition is described and evaluated. The system incorporates an auditory model front-end, an initial segmentation stage, a MLP-based phonetic segmentation and classification module, and a lexical network. The recognition is described as a search for the most likely lexical and phonetic decoding of the speech, given the acoustical observations. The search is performed by a Dynamic Programming (DP) algorithm.
Keywords: Continuous speech recognition, neural networks