Noisy speech recognition by hierarchical recurrent neural fuzzy networks

Noisy speech recognition by hierarchical recurrent neural fuzzy networks (HRNFN) is proposed. The proposed HRNFN is a hierarchical connection of two recurrent neural fuzzy networks, where one is used for noise filtering and the other for recognition. The recurrent neural fuzzy network used is the TSK-type recurrent fuzzy network (TRFN), which is constructed by recurrent fuzzy if-then rules. In n words recognition, n TRFNs are created for n words modeling. The total prediction error of each TRFN is used as recognition criterion. In filtering, n TRFNs are created, and each TRFN recognizer is connected with a corresponding TRFN filter, which filters noisy speech patterns in the feature domain before feeding them to the recognizer. Experiments on words recognition under different types of noise are performed to verify the performance of HRNFN.