A TWN Inspired Speaker Verification Processor with Hardware-friendly Weight Quantization

Speaker verification (SV) is not only a convenient biometric recognition technology but also an important method to ensure information security. Since SV systems are often deployed in mobile terminals, this places a higher demand on the trade-off between ensuring recognition accuracy and reducing system power consumption. Thus, this paper proposes an implementation of a speaker verification system based on a ternary weight network (TWN). First, we design a TWN structure for the SV system. Then a weight quantization scheme to reduce hardware storage overhead is adopted. After that, the hardware of the SV system is designed and simulated. The recognition accuracy of the proposed TWN is tested to be 83.3%@5dB, 87.9%@15dB, and 93.1%@clean, respectively. Using an industry of 22nm ULL process, the overall power consumption of the system is 16.3µW.