The invention discloses a speech emotion analysis method based on a CNN-RSC combination optimization algorithm, and belongs to the field of emotion technology. The specific steps of the method include: carrying out word segmentation processing on a corpus of a training set to obtain words; carrying out vector coding on the words; generating word vectors; using the generated word vectors as input samples to input the same into a convolutional neural network model; filtering local features among multiple words in the word vectors of the input samples by a convolutional layer to generate new feature vectors; capturing maximum values of the new feature vectors by a pooling layer to obtain feature word vector matrices; using the feature word vector matrices as feature input in recursive self-coding to obtain best vectors; adding the best vectors to an output layer for classification, and outputting emotion classification results; and carrying out error analysis and optimization processing on the emotion classification results. The combination optimization algorithm provided by the invention has a good effect on automatic learning on text emotion features, training speed of a model is also greatly improved, classification result accuracy is high, and automatic-learning ability is high.