In view of the current common speech feature extraction methods applied to the real environment, the extracted speech feature contains noise interference, which leads to the classification ambiguity in emotion recognition. Therefore, a new speech feature extraction method, namely linear prediction pitch frequency feature extraction method, is proposed. It is mainly based on linear prediction coefficient to construct a model, using the constructed model to eliminate the vocal tract response information and suppress noise interference. As this method does not achieve a better improvement for the classification ambiguity problem that occurs in emotion recognition, the LPCMCC with the same model is used to improve linear prediction pitch frequency and the comparative experiments on speech emotion recognition based on linear prediction pitch frequency, its improved features, LPCMCC and SVM are designed. The comparative experiments indicate that the average accuracy of this improved feature extraction method in the field of emotion recognition is up to 84%, which is 22% and 14% higher than that of linear pitch frequency prediction and LPCMCC, respectively. In order to test the classification effect of the improved feature in the real environment, a speech emotion recognition system based on MATLAB GUI technology is designed on the basis of the improved feature. Experimental results show that this new improved feature can effectively improve the classification ambiguity in emotion recognition, and the speech emotion system based on the improved feature can widely recognize the speaker's emotion in the presence of the noise interference.