Voice Biometrics Basics – Metrics

There are several metrics used to measure the success of a voice biometrics system:

• Enrollment rate – % of authorized users who successfully enrolled (i.e., “base” voiceprint captured)
• Verification rate - % of authorized users who are successfully verified (i.e., “sample” voice print captured and matched against “base” voiceprint)
• False acceptance rate (FAR) - % of unauthorized users who are accepted by the system. Also known as false positive or type I error.
• False rejection rate (FRR) - % of authorized users who are rejected by the system. Also known as false negative or type II error.
• Crossover error rate (CER) – point of intersection when FAR and FRR are plotted against each other (i.e., FAR = FRR). In general, the lower the CER the better.

It’s a fine art to tune a voice biometrics system to achieve simultaneously low FARs (i.e., keeping the unauthorized folks out) and low FRRs (letting the authorized folks in).

Measuring the Accuracy of Voice Biometrics

One question that frequently arises is the accuracy of voice biometrics. Organizations want to ensure that the voice biometrics application will prevent the bad guys from getting in, while at the same time, ensuring that the good guys can get in.

Basically, voice biometrics accuracy is measured by two factors:
1) False acceptance rate (FAR) – unauthorized person is accepted
2) False rejection rate (FRR) – authorized person is rejected

FAR and FRR can be plotted against each other. The point of intersection is known as the crossover error rate (CER). It should be obvious that the lower the CER the better.

It’s a fine art to tune voice biometrics applications to achieve simultaneously low FARs (i.e., keeping the unauthorized folks out) and low FRRs (letting the authorized folks in).