研究与开发采用圆周局部三值模式纹理特征的合成语音检测方法金宏辉1,简志华1,2,杨曼1,吴超1(1.杭州电子科技大学通信工程学院,浙江杭州310018;2.浙江省数据存储传输及应用技术研究重点实验室,浙江杭州310018)摘要:为了进一步提高合成语音检测的准确率,提出了一种采用圆周局部三值模式(CLTP)纹理特征的合成语音检测方法。该方法利用圆周局部三值模式提取语谱图中的纹理信息并作为语音的特征表示,采用深度残差网络作为后端分类器来判决语音真伪。实验结果表明,在ASVspoof2019数据集上,与传统的常量Q倒谱系数(CQCC)和线性预测倒谱系数(LPCC)两种特征相比,该方法在等错误率(EER)上分别降低了54.29%和2.15%,与局部三值模式(LTP)纹理特征相比,该方法在等错误率上也降低了17.14%。圆周局部三值模式由于综合考虑了邻域内中心像素与周边像素之间以及各周边像素之间的差异,更加全面地获取了语谱图的纹理信息,提高了合成语音检测的准确率。关键词:说话人验证;合成语音检测;圆周局部三值模式;深度残差网络中图分类号:TP391.42文献标志码:Adoi:10.11959/j.issn.1000−0801.2023121SyntheticspeechdetectionmethodusingtexturefeaturebasedoncircumferentiallocalternarypatternJINHonghui1,JIANZhihua1,2,YANGMan1,WUChao11.SchoolofCommunicationEngineering,HangzhouDianziUniversity,Hangzhou310018,China2.KeyLaboratoryofDataStorageandTransmissionTechnologyofZhejiangProvince,Hangzhou310018,ChinaAbstract:Inordertofurtherimprovetheaccuracyofsyntheticspeechdetection,asyntheticspeechdetectionmethodusingtexturefeaturebasedoncircumferentiallocalternarypattern(CLTP)wasproposed.ThemethodextractedthetextureinformationfromthespeechspectrogramusingtheCLTPandapplieditasthefeaturerepresentationofspeech.Thedeepresidualnetworkwasemployedastheback-endclassifiertodeterminetherealorspoofingspeech.Theex-perimentalresultsdemonstratethat,ontheASVspoof2019dataset,theproposedmethodreducestheequalerrorrate(EER)by54.29%and2.15%respectively,comparedwiththetraditionalconstantQcepstralcoefficient(CQCC)andlinearpredictivecepstralcoefficient(LPCC),andreducedtheEERby17.14%comparedwiththelocalternarypat-tern(LTP)texturefeatures.TheCLTPcomprehensivelytakesintoaccountthediffe...