ROC接受者操作特征曲线(receiver operating characteristic)常用的参数可以参考如下链接:
参考链接: https://blog.csdn.net/qq_26369907/article/details/97490991
在关键字定位(Keyword Spotting)中,引用ROC评价模型的性能,ROC纵轴为错误拒绝率,横轴为错误警告,如下图
图片来源于: SMALL-FOOTPRINT KEYWORD SPOTTING USING DEEP NEURAL NETWORKS 论文Fig. 3
而常用ROC纵轴为True Positive Rate正真阳率,横坐标False Positive Rate负真阳率,因此两者的差异,让语音识别的初学者百思不解,追本溯源发现KWS中的ROC初次出现在Google的一篇论文中,在第三章BASELINE HMM KWS SYSTEM提到:
The trade-off between false alarms (a keyword is not present but the KWS system gives a positive decision) and false rejects (a keyword is present but the KWS system gives a negative decision) is controlled by the transition probability between keyword and filler models.
其中定义:
false alarms: 关键字不存在但是KWS系统误认为存在;
false reject : 关键字存在但是KWS系统误认为不存在;
在第四章节EXPERIMENTAL RESULTS中提到:
Results are presented in the form of a modified receiver operating characteristic (ROC) curves, where we replace true positive rate with the false reject rate on Y-axis.
说白了false reject rate相当于 false negative, false alarms相当于false positive,因此FRR和FA越小越好。
ROC曲线
画ROC曲线图可以参考: https://blog.csdn.net/YE1215172385/article/details/79443552