Integrating time-delay neural networks and fuzzy HMM for phoneme-based word recognition

A speech recognition technique based on integrating time-delay neural networks and fuzzy mapping to hidden Markov models is presented. The activations from the second hidden layer of time-delay neural networks are used as outputs of a fuzzy vector quantiser (FVQ). The proposed method yields a higher recognition rate than time-delay neural networks and hidden Markov models.