Neural network noisy speech recognition and understanding with information feedback

The paper concerns noisy speech recognition and understanding. It adopts the two level modular Extended Associative Memory Neural Networks (EAMNN) for speech recognition, information feedback of the linguistic constraint reasoning and statistical inference for understanding processing. The recognition part consists of two levels: in the first level, EAMNN classifies the input data into category groups; and in the second level branch module EAMNN classifies input data into a specified category. The learning speed of two level modular EAMNNs is 9 times faster than conventional BP net, it has a high adaptive, robust fault tolerance and associative memory ability for noisy speech signals. The understanding part extracts speech recognition word candidates and predicts by inference the next word according to the statistical inference base. The linguistic rule and syntax base will reduce the candidates and acoustic recognition errors, then compare and correct error, and guide next speech processing by using information feedback, to realize recognition of sentences