CONSTRUCTION METHOD FOR DEEP LONG SHORT-TERM MEMORY RECURRENT NEURAL NETWORK ACOUSTIC MODEL BASED ON SELECTIVE ATTENTION PRINCIPLE

摘要

A construction method for a deep long short-term memory recurrent neural network acoustic model based on a selective attention principle. Change of an instant function of neurons of an auditory cortex is represented by adding an attention gate (103) unit in the deep long short-term memory recurrent neural network acoustic model, and the attention gate (103) unit is different from other gate units in that: the other gate units correspond to a time sequence on a one-to-one basis, but the attention gate (103) unit shows a short-term plasticity effect, thereby having intervals on the time sequence; extraction of robust features about Cross-talk noise and construction of a robust acoustic model are realized via the recurrent neural network acoustic model obtained by training a large amount of voice data containing the Cross-talk noise, and the purpose of increasing the robustness about the acoustic model can be achieved by restraining the influence of a non-target stream against the extraction of the features; the method can be extensively applied to the field of a plurality of machine learning related to speaker recognition and keyword recognition in voice recognition, human-machine interaction and the like.