FPGA Implementation of a Lightweight Convolutional Neural Network Classifier for Speech Emotion Recognition

This paper focuses on how to deploy CNN on FPGA for speech emotion recognition. The features of the speech data are extracted and saved to the SD card on the PC side. The trained network weight matrix is also held on the SD card. We designed the modules of CNN by HLS. FPGA reads the speech features and the network weight matrix from the SD card; then, the deployed CNN model performs the emotion prediction. The results are sent by UART serial port to the PC side. In this paper, the Berlin EmoDB corpus is used for the experiments. We first experiment on the software side and achieved an 85.67% recognition accuracy. Then we get the same accuracy on FPGA verification.