Dynamic programming prediction errors of recurrent neural fuzzy networks for speech recognition

This paper proposes Mandarin phrase recognition using dynamic programming (DP) prediction errors of singleton-type recurrent neural fuzzy networks (SRNFNs). This method is called DP-SRNFN. The recurrent property of SRNFN makes it suitable for processing temporal speech patterns. A Mandarin phrase comprises monosyllabic words. SRNFN training is based on the word unit. There are Nw SRNFNs for modeling Nw words, and each SRNFN receives the current frame feature and predicts the next one of its modeling word. In recognizing NPNP phrases, the prediction error of each trained SRNFN is computed, and DP is used to find the optimal path that maps the input frames to the best matched SRNFNs (words) for each of the NPNP phrases. The accumulated error of each phrase model is computed from its optimal path and the one with the minimum error is the recognition result. To verify DP-SRNFN performance, this study conducted experiments on recognizing 30 Mandarin phrases. SRNFN training with noisy features for phrase recognition under different noisy environments was also conducted. DP-SRNFN performance is compared with the hidden Markov models (HMMs). Results show that DP-SRNFN achieves higher recognition rates than HMM in both clean and noisy environments.