False acceptance rate (FAR) is how often a biometric system wrongly accepts an impostor, and false rejection rate (FRR) is how often it wrongly rejects the genuine user. They trade off against each other: a stricter threshold lowers FAR but raises FRR, and a looser threshold does the reverse. The right balance depends on whether security or convenience matters more.
If you are evaluating or tuning a voice biometric system, two metrics come up constantly: false acceptance rate and false rejection rate. They are the core measures of accuracy, and they pull in opposite directions. Understanding the trade-off between them is essential to setting a sensible threshold and comparing systems fairly. This guide explains both and how to balance them.
What is false acceptance rate (FAR)?
False acceptance rate is the proportion of impostor attempts that the system wrongly accepts as genuine. It is a security metric: a high FAR means impostors get through too often. A false acceptance is the dangerous error, because it grants access to the wrong person. Lower FAR means stronger security.
What is false rejection rate (FRR)?
False rejection rate is the proportion of genuine attempts that the system wrongly rejects. It is a convenience metric: a high FRR frustrates real users who are denied access and must retry or seek help. A false rejection is the inconvenient error. Lower FRR means a smoother user experience.
FAR vs. FRR: the trade-off
The two rates are linked through the matching threshold, and improving one usually worsens the other.
| Threshold | Effect on FAR | Effect on FRR |
|---|---|---|
| Stricter (higher) | Lower (more secure) | Higher (more rejections) |
| Looser (lower) | Higher (less secure) | Lower (fewer rejections) |
Raising the threshold demands a closer match, which keeps more impostors out but also rejects more genuine users whose sample varies slightly. Lowering it accepts looser matches, letting more genuine users in but also more impostors. There is no single setting that minimizes both at once.
Security versus convenience
Because of this trade-off, the right balance depends on context. For high-value actions such as large transfers, you weight security and accept a lower FAR at the cost of a higher FRR. For low-risk, high-volume interactions, you may favor convenience with a lower FRR. Many systems use different thresholds for different actions, tightening for sensitive operations.
How to balance FAR and FRR
- Set thresholds by risk. Use stricter thresholds for high-value actions and looser ones for low-risk steps.
- Use a review band. Route borderline scores to a help desk or second factor rather than forcing a binary decision.
- Compare systems at the same operating point. Or use the equal error rate, where FAR and FRR are equal, for a single comparison number.
- Do not forget spoofing. Deepfakes attack FAR specifically, so pair accuracy with anti-spoofing.
Why anti-spoofing matters for FAR
A convincing deepfake is an impostor attempt designed to be accepted, so it directly targets FAR. A system with a good FAR against ordinary impostors can still be vulnerable to clones without anti-spoofing. This is why VoiceVantage pairs accuracy with proven deepfake detection: it blocked all five leading deepfake tools in 533 independent tests, keeping false acceptance low even against synthetic voices.
Frequently Asked Questions
What is the difference between FAR and FRR?
FAR is how often the system wrongly accepts an impostor (a security error), while FRR is how often it wrongly rejects the genuine user (a convenience error).
Which is worse, a false acceptance or a false rejection?
A false acceptance is the more dangerous error because it grants access to the wrong person; a false rejection is inconvenient but not a security breach.
How do I lower FAR?
Raise the matching threshold, which reduces false acceptances but increases false rejections, so balance it against user experience.
Can I minimize both FAR and FRR at once?
Not with a single threshold, since they trade off. You can improve both overall by using a more accurate engine or a review band for borderline scores.
How does a deepfake affect FAR?
A deepfake is an impostor attempt aimed at being accepted, so it targets FAR directly, which is why anti-spoofing is essential.
What threshold should I use?
Set it by risk: stricter for high-value actions to lower FAR, looser for low-risk actions to lower FRR, with a fallback for borderline cases.
Need low false acceptance even against deepfakes? See how VoiceVantage performs in independent testing.