基于深度学习的声纹识别身份验证系统设计

Deep Learning Based Speaker Verification Identity Authentication System Design

该文设计了一种将声纹识别技术部署于AI推理计算设备上的身份验证系统,主要包含待验证音频采集模块和音频特征向量提取模块等部分。基于RK3568高性能芯片的主控制单元通过音频采集模块进行待验证音频采集,之后将待验证音频进行预处理获取FBank特征谱。预处理后的音频数据将通过ECAPA-TDNN声纹识别模型进行嵌入特征向量提取,从而获得具有辨识度的嵌入特征向量。嵌入特征向量经过在已注册特征数据库中搜寻并计算两者之间的余弦相似度,通过与阈值进行比较,从而获得待验证人员的身份信息。最后,根据预设的身份权限信息,执行相对应的操作。通过实验验证,当余弦相似度阈值设置为0.3时,能够获得很好的验证效果,因此证明了该系统在实际部署的可行性。

The paper designs an identity verification system that deploys speaker verification technology on an AI inference computing device,which mainly contains the parts of audio to be verified acquisition module and audio feature vector extraction module.The main control unit based on the RK3568 high-performance chip carries out the audio to be verified audio acquisition through the audio acquisition module,after which the audio to be verified is preprocessed 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 recognizable degrees.The embedded feature vector is searched in the registered feature database and the cosine similarity 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 information.It is verified through experimental tests that when the cosine similarity threshold is set to 0.3,good verification results can be obtained,proving this system design’s feasibility in practical deployment.