Selective attention for noise robust speech recognition
K.Y. Park, Soo-Young Lee · 1999
Based on biological selective attention mechanism, a new algorithm was developed to improve recognition accuracy of isolated speeches in noisy environments. The attenuating “early filtering” model was implemented by inserting an attention layer just after the input layer. Attention gains, i.e., one-to-one synaptic weights between the original sensor input vector and the attended input vector at the attention layer, were adapted with the error backpropagation algorithm. After attention adaptation, the distance between the original sensor input and the attended vectors became a confidence measure for classification. The developed algorithm demonstrated high recognition rates for isolated Korean words in noisy environments.