Equal error rate (EER) is the point at which a biometric system’s false acceptance rate equals its false rejection rate. It gives a single number to summarize and compare accuracy: a lower EER means a more accurate system. EER is useful for quick comparison, but it reflects one operating point and does not measure a system’s resistance to deepfakes.
When comparing voice biometric systems, you will often see accuracy quoted as a single percentage called the equal error rate. It is a convenient summary, but only if you understand what it does and does not tell you. This guide explains EER, why it is used, how it relates to FAR and FRR, and its limitations.
What is equal error rate (EER)?
Equal error rate is the point at which the false acceptance rate and the false rejection rate are equal. As you adjust a system’s threshold, FAR and FRR move in opposite directions; at one specific threshold they cross and become equal, and that shared value is the EER. It is expressed as a percentage, and a lower EER indicates a more accurate system overall.
Why is EER used?
EER exists because FAR and FRR alone are hard to compare across systems, since each depends on the chosen threshold. By fixing the comparison at the point where the two errors are equal, EER gives a single, threshold-independent number. That makes it a quick way to rank systems: between two engines, the one with the lower EER is generally more accurate.
How is EER determined?
EER is found by testing a system across a range of thresholds and measuring FAR and FRR at each. Plotting the two against each other traces a curve (often called a detection error trade-off curve), and the EER is the point where FAR equals FRR. It is a summary of test results, not a live setting the system runs at; in production you tune the threshold to favor security or convenience as needed.
What is a good EER?
Lower is better, and leading voice biometric engines achieve low single-digit or fractional EERs under good conditions. However, an EER is only meaningful alongside the conditions it was measured in: audio quality, sample length, and the test population all affect it. A very low EER measured in ideal lab conditions may not hold on noisy real-world channels, so always ask how it was measured.
The limitations of EER
- It reflects one operating point. Real systems rarely run at the equal-error threshold; you tune for security or convenience, so live FAR and FRR differ from the EER.
- It hides the trade-off. Two systems with the same EER can behave differently at the threshold you actually use.
- It does not measure spoofing. EER is measured against ordinary impostors, not deepfakes, so a low EER says nothing about anti-spoofing.
- Conditions matter. An EER is only comparable when measured under similar conditions.
EER and anti-spoofing are different
This is the most important caveat. A system can have an excellent EER and still be fooled by a cloned voice, because EER does not test synthetic attacks. Deepfake resistance is measured separately, through testing against real deepfake tools. This is why VoiceVantage points to its 533 independent deepfake tests, in which it blocked all five leading deepfake tools, rather than relying on a single accuracy number. When comparing vendors, look at both accuracy metrics like EER and independent anti-spoofing evidence.
Using EER to compare vendors
Treat EER as one input, not the whole decision. Ask each vendor for their EER and the conditions it was measured under, run your own proof of concept, and separately require independent deepfake test evidence. Combined, these give a fuller picture than any single figure. See our buyer’s comparison of voice biometrics vendors for a full framework.
Frequently Asked Questions
What does equal error rate (EER) mean?
It is the point where a system’s false acceptance rate equals its false rejection rate, giving a single number to summarize accuracy. Lower is better.
Is a lower EER better?
Yes. A lower EER indicates a more accurate system overall, though the conditions it was measured under matter.
Does a system run at its EER threshold?
Usually not. EER is a summary of test results; in production you tune the threshold to favor security or convenience.
Does EER measure deepfake resistance?
No. EER is measured against ordinary impostors, not synthetic voices, so anti-spoofing must be evaluated separately.
What is a good EER for voice biometrics?
Lower single-digit or fractional EERs are strong under good conditions, but always check how the figure was measured.
How should I use EER when comparing vendors?
As one input alongside your own proof of concept and independent deepfake test evidence, not as the sole basis for a decision.
Comparing accuracy and anti-spoofing? Ask VoiceVantage for both its metrics and independent test results.