instead of 43 to verify identity
Biometrics replaced the security questions: the customer simply talks and the agent sees “identity confirmed”.
A bank’s contact centreVoice Biometrics · 2025Client under NDA
The customer confirms their identity by voice
The check takes a few seconds of ordinary speech: no passphrases and no security questions.
Three ways to be sure it is the right person calling
One and the same question — “is this really you?” — and three answers to it.
| Three ways to be sure it is the right person calling | Security questions | Passport details | Voice biometrics |
|---|---|---|---|
| How long it takes | 43 seconds | 40–60 seconds | 5–7 seconds |
| What can be stolen | The answers can be overheard | Copies of a passport travel from hand to hand | A voiceprint is useless without the voice |
| What can be forgotten | Your mother’s maiden name | The passport within reach | Nothing |
| What the customer does | Remembers and answers | Finds the passport, reads out the series and number | Simply talks |
Both phases happen in the background of an ordinary conversation
The customer performs no exercises and reads out no control phrases. They simply talk to the agent.
Enrolment
Once, on the first contact
A voiceprint is built from natural speech: a set of numbers by which the voice is recognised later. The recording itself need not be kept.
Verification
On every following call
The system matches the voice against the voiceprint and returns a match score. The agent sees the result before they even get to the question.
It does not matter which language the customer speaks, or what they say
What is checked is how the voice sounds, not the sentence spoken. That is why there are no control words to overhear.
Uzbek
Assalomu alaykum, kartam boʻyicha savolim bor
Recognised by voice
Russian
Здравствуйте, у меня вопрос по карте
Recognised by voice
English
Hello, I have a question about my card
Recognised by voice
The same person, three languages, one and the same voiceprint.
And the reverse task: recognising someone who would rather not be
A voice that has been caught once goes into the database: next time the alarm is raised in the first seconds.
A database of fraudsters’ voices
Your own for your organisation, plus an industry-wide one if you take part in it.
Archive search
Find every past call made by this voice — even if it gave different names.
Security alert
The security team gets the signal during the call, not a week later.
AlertThe voice matched a record in the fraudster database
Find every call made by one person
A scenario for security teams and public bodies: we take a voice sample and find every recording it appears in, in a fraction of a second.
- You upload a fragment of a recording with the voice in question
- The system builds a voiceprint from it
- It returns a list of the calls where that voice appears
01A fragment of a recordingAny conversation where the voice in question is heard.
02The voiceprintA set of numbers: speech cannot be restored from it.
Two checks from one voiceprint
Match against the referenceWhether this is the person they claim to be.
Match against the watchlistWhether the voice has called before under another name.
An honest word about accuracy
over 98%of matches on telephone recordingsAccuracy is not one number but a balance of two errors. The threshold is tuned to your risk.
False match
The system took a stranger for your customer. This is the dangerous one: it is exactly what fraudsters look for. For money operations the threshold is raised.
False rejection
The system did not recognise your own customer — they had a cold, for instance. Not a problem: the agent falls back to the usual check.
Consent, storage and the law
Customer consent
The customer consents to the processing of biometric data before the voiceprint is created. The wording and the way it is recorded are agreed with your lawyer.
What exactly is stored
A voiceprint is a set of numbers; a voice recording cannot be restored from it. Keeping the original recording is not required.
Where it is stored
In the on-premise option voiceprints never leave your network. That is usually what the Uzbek personal data law requires as well.
Withdrawing consent
The customer may withdraw consent, the voiceprint is deleted and verification falls back to the usual scenario.
Where this is already needed
Banking
A check before an account operation, protection against social engineering.
Insurance
Taking claims over the phone and spotting repeat fraudulent enquiries.
Public services
Citizen enquiries without a visit in person, working inside a closed network.
Telecom
Changing a tariff and restoring a SIM card without going to an office.
What changes when the customer is recognised by voice
The check stops being an interrogation at the start of every conversation.
- No recalling a passphrase or a mother’s maiden name — talking is enough.
- The conversation starts with the matter at hand, not a minute of formalities.
- They are recognised from someone else’s number too: the voice belongs to the person, not the SIM.
- The check runs in the background while the customer explains the task.
- Every conversation is shorter — and that shows across the whole shift, not in one call.
- The agent does not decide “close enough or not”: the system does.
- One voice introducing itself as different people shows up immediately.
- Known fraudsters go onto a list: one attempt links all the others.
- The alert reaches the security team during the call, not after the review.
- The decision rests on the recording and the match, not on the agent’s memory.
- A voiceprint cannot be turned back into speech: this is not storing a recording.
- Customer consent and retention periods are set by policy, not by habit.
How it is rolled out
At first biometrics works silently and decides nothing — it only learns and compares.
Collecting voiceprints
A voiceprint is recorded from an ordinary conversation with the customer’s consent. Nothing has to be dictated or read out.
Observation mode
The system matches voices and shows the result, but the decision stays with the agent. This is how you see the accuracy on your own calls.
Connecting it to the process
The verification result arrives in the agent’s record and in your system. We agree what happens on a match and on a doubt.
Adding fraud protection
We set up the database of known voices and configure alerts for the security team.
Still have questions?
Can the system be fooled with a voice recording?
A liveness check works against recordings: the system tells live speech from playback by features a recording does not have. For high-risk operations it is switched on as a rule and the match threshold is raised.
What if the customer has a cold or sounds hoarse?
The match will come out lower than usual and the system will say so. That is a false rejection: the agent simply falls back to the usual check. The voiceprint is not spoiled by it.
Where are the voiceprints stored?
Wherever you decide. In the on-premise option voiceprints stay inside your network and never leave it. A voiceprint is a set of numbers; the voice cannot be restored from it.
Does it work on a poor line?
Yes, the models are trained on telephone recordings with noise and compression. On a really bad line the system asks for a few more seconds of speech rather than giving an unreliable answer.
Let us check the accuracy on your own recordings
Send us a set of calls with known speakers. We will show how the system tells them apart and tune the threshold to your risk.
