Note
Access to this page requires authorization. You can try signing in or changing directories.
Access to this page requires authorization. You can try changing directories.
[This article is prerelease documentation and is subject to change.]
Important
You're responsible for configuring and using this Biometric authentication and enrollment functionality in compliance with applicable law, including but not limited to biometric and privacy laws (for example, BIPA and the GDPR), including: (i) providing notice to data subjects, including with respect to retention periods and destruction; (ii) obtaining legally sufficient consent from data subjects; and (iii) processing and deleting the biometric data and biometric identifiers. Biometric data and biometric identifiers will have the meanings set forth in applicable biometric and privacy laws and include equivalent terms set forth in any applicable law. You will defend and indemnify Microsoft against any claims made by a third party related to this biometric functionality, including but not limited to the adequacy of consent.
Biometric authentication verifies a caller's identity from the unique characteristics of their speech. In Contact Center Intelligence, the Biometric authentication service uses voice biometrics to confirm that a caller is who they claim to be and to detect attempts to imitate or replay a voice. Understanding the underlying concepts helps you configure authentication behavior and interpret authentication outcomes with confidence.
Important
- This is a preview feature.
- Preview features aren’t meant for production use and might have restricted functionality. These features are subject to supplemental terms of use, and are available before an official release so that customers can get early access and provide feedback.
What a voiceprint is
A voiceprint is a secure mathematical representation of the distinctive characteristics of a customer's speech, associated with a customer identifier. It's a mathematical model of the customer's speech rather than the audio itself. Voiceprints are stored within your Dataverse environment.
Your customers' voiceprint is created during enrollment and is the reference that biometric authentication compares against on later calls.
How authentication works
During authentication, biometric authentication evaluates two things together before it returns a result:
- Speaker similarity — whether the voice on the current call matches the enrolled voiceprint for the claimed identity.
- Anti-spoofing signals — whether the incoming speech contains audio artifacts indicating potential synthetic speech or a recording of a live person..
By combining a similarity assessment with anti-spoofing signals, biometric authentication verifies both that the voice matches and that it's a genuine, live speaker. The service returns a real-time result that downstream experiences can act on.
How performance is measured
Voice biometrics technology used in Biometric authentication is probabilistic: a person's voice varies with time of day, stress, illness, environment, and the device they use. The decisioning engine handles this variation well, but occasional incorrect decisions are normal and expected. Two rates describe the trade-off, and they move in opposite directions:
- False accept rate (FA) — the percentage of imposters who receive a positive authentication decision.
- False reject rate (FR) — the percentage of genuine customers who receive a negative decision.
Raising the voice-authentication threshold lowers the false accept rate but raises the false reject rate. You tune this balance to match your organization's risk tolerance.
External factors that affect performance
Some factors that influence decisioning performance are outside the configuration:
- Audio quality — telephony platform, network carrier, geographic region, and device all matter. A low-quality VoIP call performs worse than a strong Wi-Fi or 5G call.
- Caller behavior — speakerphone or Bluetooth use and distance from the microphone reduce quality compared with speaking directly into a phone.
Recorded audio and non-natural speech aren't always fraudulent. Hold music, automated welcome messages, IVR prompts, and background media appear legitimately in up to 10% of contact center calls, so some authentication failures from such audio are normal and expected.
Anti-spoofing and synthetic-speech detection
Anti-spoofing protects against attempts to defeat voice authentication by presenting speech that didn't come from the live, enrolled customer. Like voice biometrics, both synthetic speech and playback detection are probabilistic decisioning technologies. Biometric authentication anti-spoofing covers two categories:
- Synthetic (non-natural) speech — real-time detection of artificially generated speech, such as text-to-speech or voice-cloned audio.
- Playback and recordings — detection of previously captured audio that's replayed during the call.
These checks run alongside the speaker-similarity evaluation so that a close voice match alone isn't enough to authenticate when the speech appears to be synthetic or replayed.
Net speech and duration guidance
Only the customer's actual speech — called net speech — counts toward enrollment and authentication. Silence, pauses, hold time, and non-speech audio don't count. Two people talking for a minute might yield only a few seconds of net speech per person.
As general guidance:
- Enrollment needs about 6–12 seconds of net speech to build a voiceprint.
- Authentication needs about 3–6 seconds of net speech to evaluate a match.
The exact amount of net speech required is configurable, so you can tune it to balance accuracy against how much speech a customer must provide.
Longer net speech generally improves authentication accuracy and strengthens playback and synthetic-speech detection. Less net speech increases the chance that authentication fails or can't complete. This trade-off is why voice-agent prompts and conversation design aim to collect enough natural speech without adding friction.
Why biometric authentication matters
Biometric authentication changes how a contact center verifies identity:
- Reduced friction — customers are verified from their natural speech instead of answering knowledge-based questions or entering one-time passcodes.
- An added authentication factor — voice provides another factor that can complement or replace traditional verification methods.
- Spoof protection — real-time anti-spoofing helps defend against synthetic speech and recordings, which knowledge-based checks can't detect.
Together, these benefits let a contact center authenticate more callers with less effort while raising the bar against voice-based fraud.
Interaction types
Biometric authentication applies across two interaction types, which differ in how the customer's speech is collected:
- Human-to-Bot (H2B) — Authentication happens while the customer interacts with a voice agent. As the voice agent collects the customer's speech, biometric authentication uses the same audio to enroll or authenticate. If the customer doesn't provide enough speech, the voice agent can ask more questions to gather it, and biometric authentication returns the result to the agent.
- Human-to-Human (H2H) — Authentication happens passively during a natural conversation between the customer and a customer service representative (CSR).
The same underlying voiceprint, speaker-similarity, and anti-spoofing concepts apply to both interaction types. Only the experience that gathers the speech differs, not the biometric evaluation itself.