Abstract: MFCC（Mel Frequence Cepstral Coefficients） only reflects speech static feature so its dialect recognition rate is low, while SDC（Shifted Delta Cepstra） reflects speech dynamic feature because of considering the connections between several speech frames. For combination of static and dynamic features, MFCC and SDC extracted from Mandarin, Shanghai dialect, Cantonese, Minnan dialect are employed as the feature vector with SVM（Support Vector Machine） for the dialect identification, and effects on performance of different parameters for SDC are studied. Simulation results demonstrate that the dialect recognition rate with static and dynamic features can be up to 92.5%, but its increase is based on the cost of the working time.

Key words: dialect identification, Mel Frequence Cepstral Coefficients（MFCC）, Shifted Delta Cepstra（SDC）, Support Vector Machine（SVM）

摘要： 针对MFCC仅反映语音静态特征导致的方言识别率低的问题，而SDC由于考虑了前后帧差分倒谱的影响，能反映语音的动态特征；同时考虑方言的静态与动态特征，对普通话、上海话、广东话和闽南话4种方言进行MFCC特征和SDC特征提取，将其两组特征组合送入支持向量机进行辨识，并研究了针对4种方言的SDC的局部最优参数组合。仿真实验结果表明，同时考虑方言的静态与动态特征方法的识别率高达92.5%，但识别率的提高是以延长运算时间为代价的。

关键词: 方言辨识, Mel频率倒谱系数, 滑动差分倒谱特征, 支持向量机