This work presents an in-depth study on the classification of regional accents in Mandarin speech. Mel-Frequency Cepstral Coefficients (MFCC) with 13 features are used as the input in this work. The dataset used is generated by speakers from 96 cities, which covers 13 large dialect areas in China. Eight dialect areas which cover about 75% of the dataset are selected in this work. The work explores 1-Dimensional Convolutional Neural Network with Stochastic Gradient Descent as an optimizer to extract the acoustic feature and get relative high accuracy of 67.15%. It is an important step to reduce the character error rate of ASR models. Meanwhile, it is useful to narrow a criminal's location of living.