Speech emotion recognition(SER)is the use of speech signals to estimate the state of emotion.At present,machine learning is one of the main research methods of SER,the test and training dataS of tradition-al machine learning all have the same distribution and feature space,but the data ofspeech is accessed from dif-ferent environments and devices,with different distribu-tion characteristics in real life.Thus,the traditional ma-chine learning method is applied to the poor performance of SER.This paper proposes a multi-distributed SER method based on Mel frequency cepstogram(MFCC)and parameter transfer.The method is based on single-layer long short-term memory(LSTM),pre-trained inception-v3 network and multi-distribution corpus.The speech pre-processed MFCC is taken as the input of single-layer LSTM,and input to the pre-trained inception-v3 network.The features are extracted through the pre-trained incep-tion-v3 model.Then the features are sent to the newly defined the fully connected layer and classification layer,let the parameters of the fully connected layer be fine-tuned,finally get the classification result.The experi-ment proves that the method can effectively complete the classification of multi-distribution speech emotions and is more effective than the traditional machine learning framework of SER.