Lightweight Implementation of FPGA-Based Environmental Sound Recognition System

The intensive computation of existing sound feature extraction and recognition algorithms makes it difficult to work on edge computing devices with limited computation power. To address this issue, this research designs implementation of sound feature extraction and recognition systems on field programmable gate array (FPGA) in real-time applications. Since the sound feature extraction and recognition algorithms are quite computationally intensive, this paper first optimizes the sound Mel frequency spectrum coefficients feature extraction algorithm and the convolutional neural network recognition algorithm. After the optimization, the data volume of the sound extraction algorithm is reduced by up to 83.94%, and the parameter volume of the recognition algorithm is reduced by up to 65.63%, making it feasible to be implemented on FPGA. Then the real-time sound recognition system implemented on FPGA can reach a recognition accuracy of 88.33% in real-world environments.