Design and implementation of a voiceprint recognition system based on FPGA

A human voiceprint recognition system has been constructed based on a field programmable gate array (FPGA) platform. By processing the original voice signal and combining it with an optimized deep neural network architecture, the system achieves rapid extraction and precise matching of voiceprint features. By leverages the parallel processing advantages of FPGA, mapping complex algorithms onto hardware circuit structures, data processing speed was significantly improved with high accuracy of the recognition. The results indicate that the FPGA-based voice recognition system has achieved a significant increase in data processing rates while maintaining high recognition accuracy. Deep feature extraction architecture combining multi-branch pyramid convolution (MBPC) and adaptive spectral normalization (ASNorm) reduces the equal error rate (EER) to 2.15% on the VOiCES-2019 test set, improving robustness by 37.4% over the baseline. A heterogeneous precision mapping strategy and spatiotemporal scheduling engine achieve 172ms latency and 3.2 TOPS/W energy efficiency on Xilinx UltraScale+ platforms, an 8.7× improvement over GPUs. A dynamic threshold decision circuit reduces the false acceptance rate (FAR) by 2.1 orders of magnitude under 0dB SNR via error feedback, enhancing the stability in complex acoustic environments.