Speaker classification based on combined neural network and fuzzy decision

Speech recognition has played a key role in man-machine communication. The authors present a new speaker classifier based on combined neural network and integrate the fuzzy decision and knowledge supervision. Here, they first deal with context-dependent speaker classification. The input speech is sampled at 10 kHz and a 12-order cepstrum is obtained in real-time by a TMS320C25 board. A 2-stage combined neural network is proposed. The first stage is the neural prediction network in order to include the time correlation of the speech signal and to compress and normalize the parameters. The authors use the back-and-forward prediction network. After convergence, the speaker information is embedded in the weights of the predictor. The authors modify the conventional BP algorithm to the fuzzy BP algorithm for training, which has obtained better results. Considering its outstanding ability of characteristic extraction and auto-clustering, the main network is the Self-organizing Feature Map (SOFM). The weights of the predictor become the input of the SOFM which naturally needs no normalization of its input parameters. Using the algorithm proposed by T. Kohonen, the map was rapidly trained and afterwards it is marked for each speaker who says the same word. Then, the characteristics of each speaker are reflected in the map.