EmotionEdge: An Efficient Framework for Speech Emotion Recognition

Emotion recognition is an essential task in artificial intelligence applications. Human-centered smart systems with affective interaction can provide services of higher quality, especially speech emotion recognition. Considering the end device does not have enough memory and computational resources available, it will introduce a communication delay if it offloads the processing task to a cloud. We propose a speech emotion recognition system assisted by edge servers and cloud computing. In this system, feature extraction and classification are conducted on end devices. The databases are stored on the cloud and large-scale feature extraction is executed on the cloud. Model training is implemented on edge servers. Differing from previous work, a search block of approximate nearest neighbors based on product quantization is added as the source of training data to further reduce the amount of calculation. With a small-scale training set, the SVMs are employed as classifiers due to the high quality of SVM on a small dataset. Experiments show that our solution is a potential solution for emotion detection with limited resources.