Deep Learning Based Speaker Verification Identity Authentication System Design
The paper designs an identity verification system that deploys speaker verification technology on an Al in-ference computing device,which mainly contains the parts of audio to be verified acquisition module and audio fea-ture vector extraction module.The main control unit based on the RK3568 high-performance chip carries out the au-dio to be verified audio acquisition through the audio acquisition module,after which the audio to be verified is pre-processed to obtain the FBank feature spectrum.The preprocessed audio data will be subjected to embedded feature vector extraction through the ECAPA-TDNN speaker verification model to obtain embedded feature vectors with rec-ognizable degrees.The embedded feature vector is searched in the registered feature database and the cosine similar-ity between them is calculated,and the identity information of the person to be verified is obtained by comparing it with the threshold value.Finally,the corresponding operation is executed according to the preset identity authority in-formation.It is verified through experimental tests that when the cosine similarity threshold is set to 0.3,good verifi-cation results can be obtained,proving this system design's feasibility in practical deployment.
speaker verificationRK3568 chipdeep learningAI model deployment
NETL
NSTL
万方数据
