一种基于预训练模型的语音深度伪造算法识别方法  

A Pre-trained Model Based Recognition Method for Speech Deepfake Algorithms

作　　者：田野 罗曦 许斌 葛珊[1] 张向阳[1] TIAN Ye;LUO Xi;XU Bin;GE Shan;ZHANG Xiangyang(The Third Research Institute of China Electronics Technology Group Corporation,Beijing 100015,China)

机构地区：[1]中国电子科技集团公司第三研究所,北京100015

出　　处：《电声技术》2024年第2期28-31,35,共5页Audio Engineering

摘　　要：为提高语音深度伪造算法识别模型的准确性和对未知伪造算法识别的泛化性,文章提出一种基于预训练模型的识别方法。基于真伪语音数据集,微调训练HuBERT预训练模型,并基于模型输出的深层嵌入特征构建流形空间,通过度量不同伪造算法下语音数据流形空间的测地线距离进行伪造算法的判定。实验表明,所提方法可以较为有效地实现对已知和未知伪造算法的识别。To improve the accuracy of the recognition model of speech deepfake algorithms and the generalization of the recognition of unknown deepfake algorithms,a recognition method based on the pre-trained model is proposed.Based on the real and fake speech dataset,the HuBERT pre-trained model is fine-tuned and the manifold space is constructed based on the deep embedded features output from the model,and the determination of deepfake algorithms is carried out by measuring the geodesic distances between the manifold spaces of different deepfake algorithms.Experiments show that the proposed method can realize the recognition of known and unknown deepfake algorithms more effectively.

关 键 词：深度伪造算法识别 预训练模型 流形测度 

分 类 号：TB54[理学—物理]