Implementation of a CRNN-based low-power keyword recognition system on FPGA

Abstract

A low-power and high-precision reconfigurable processor based on optimized convolutional recurrent neural network is proposed for noise robust keyword recognition. In order to create a low-power and high-precision system, we implemented a reconfigurable CRNN and quantization network on FPGA, which greatly reduced the use of DSP, BRAM, LUT and other resources. Our system can identify some keywords, such as "yes", "no", "down" and "up" within 50ms, and at a signal-to-noise ratio of-5dB, the actual accuracy reaches 86.4%.