In order to evaluate a voiceprint recognition model, we need compute eer metric. In this tutorial, we will introduce some metrics TPR, FPR, FAR, FRR to help you understand how to compute ERR.
TPR and FPR
TPR and FPR can be computed as follows:
TPR = TP / (TP + FN)
FPR = FP / (FP + TN)
You can learn more on these two metrics in here:
Understand TPR, FPR, Precision and Recall Metrics in Machine Learning – Machine Learning Tutorial
FAR and FRR
This tutorial gives us some basic introduction on them.
Compute FAR, FRR and EER Metrics in TensorFlow – TensorFlow Tutorial
Here:
FAR = False Acceptance Rate
FRR = False Rejection Rate
EER =FAR = FRR
From paper: Robust Performance Metrics for Authentication Systems, we can find:
FPR is sometimes called the false accept rate (FAR). It means FPR = FAR
The false negative rate (FNR) which is alternatively called the false reject rate (FRR). It means FNR = FRR
FNR = 1-TPR = FRR
How to compute EER?
When FAR = FRR, we can get EER. Moreover, we can compute it by ROC curve.
ROC curve looks like:
In ROC curve, the x axis is FPR (FAR), the y axis is TPR. The line from (0, 1) and (1.0) is FRR.
Then we can compute EER easily.