Abstract: Based on the modified QuickNet software, we proposed a supervised DNN layerwise pre-training method for dialect speech recognition. The pre-training will start from a 3-layer neural network till the maximum layer, during which we will do supervised training. The initial weights of a new layer are composed of the partial trained weights of lower level network and the randomized weights closed to the output layer. Then we will do traditional back-propagation training when the initial weights of the maximum layer network are obtained. This method achieved a relatively higher recognition rate compared with normal neural network training and can be used in mobile speech recognition apps, the recognition of dialects speech and so on.

Key words: deep neural network, dialects speech recognition, QuickNet