AIoT Smart ICs & Systems Lab
Linrui designed a speaker recognition algorithm with the time delay neural network (TDNN). By setting up a rational residual network structure and applying a comprehensive speech dataset, the module’s equal error rate (EER) is reduced from 14.22% to 0.196% when full-time enrollment and verification.
He also made the algorithms hardware-friendly, transferring long audio segments into shorter segments and integrating the processing results from short audio segments. Achieved just about 0.05% performance deterioration on the systems. (0.245% compared with 0.196%)
