Out-of-Vocabulary Rejection based on Selective Attention Model

摘要

Based on a psychological selective attention theory a new algorithm is developed to provide reliable out-of-vocabulary (OOV) rejection for speech recognition systems in noisy environments. The developed attention model is based on Broadbent's ‘early filtering’ theory, and the attention adaptation process utilizes a gradient-descent error minimization algorithm with error backpropagation rule. The developed model is applied to isolated-word recognition tasks, and much higher in-vocabulary recognition rates are achieved with the same out-of-vocabulary rejection rates.