基于知识蒸馏的说话人验证模型轻量化方案.pdfVIP

-S系统程A电声技戴

OystemEngineeringudioEngineering

文献引用格式：钱建宇，基于知识蒸馏的说话人验证模型轻量化方案[].电声技术，2024，48（7)：28-31，42.

QIANJY.Aknowledgedistillationbasedmodeloptimizationmethodforspeakerverification[JJ.AudioEngineering2024,48

(7):28-31,42.

中图分类号：TP183文献标识码：AD0I:10.16311/j.audioe.2024.07.009

基于知识蒸馏的说话人验证模型轻量化方案

钱建宇

（上海大学通信与信息工程学院，上海200444)

摘要：利用大规模无监督数据训练的预训练模型具有非常好的泛化性，只需在小规模标注数据上进行微调，就能在相应任务

上有所提高。然而，预训练模型加上下游模型通常拥有较大的计算量和较慢的推理速度，不适合在低性能的边缘设备上部署，

也难以满足需要实时化任务的场景。基于此，提出一种基于知识蒸馏的说话人验证模型轻量化方案，通过将预训练模型和下

游模型蒸馏到一个学生网络上，实现整个任务流程的轻量化。

关键词：说话人验证；模型轻量化；知识蒸馏

AKnowledgeDistillationBasedModelOptimizationMethodforSpeakerVerification

QIANJianyu

(SchoolofCommunicationandInformationEngineering,ShanghaiUniversity,Shanghai200444,China)

Abstract:Thepretrainedmodeltrainedonlarge-scaleunsuperviseddatahasexcellentgeneralizationability,andcanbeimproved

incorrespondingtasksbyfine-tuningonsmall-scaleannotateddata.However,pretrainedmodelscombinedwithupstreamand

downstreammodelsoftenhavealargecomputationalloadandslowinferencespeed,makingthemunsuitablefordeploymentonlow

performanceedgedevicesanddifficulttomeetscenariosthatrequirereal-timetasks.Basedonthis,aoptimizationmethodforspeaker

verificationmodelbasedonknowledgedistillationisproposed,whichachievesthelightweightingoftheentiretaskprocessbydistilling

thepretrainedmodelanddownstreammodelontoastudentnetwork.

Keywords:speakerverification;modellightweighted;knowledgedistilla