AI use case

AI voice biometric caller authentication with synthetic voice detection

AI voice biometrics compares a caller's voice with a voiceprint the customer enrolled earlier, in the first seconds of a call, so a genuine customer can skip the security questions and a likely impostor is flagged to an agent. Many systems also check whether the voice is live, a recording or synthetic.

Last verified 18 November 2026 · 3 public deployments

At least 4.2 million
Users served
Australian Taxation Office (independent claim).
USD 66,667 to USD 750,000
Indicative value per year
A retail bank with 2 million inbound service calls a year. Worked example, see how it is calculated.

What problem does it solve?

Every call to a bank, a tax office or a telecom operator starts with the same step: the agent has to find out who is on the line. Knowledge based questions (date of birth, address, the last transaction) take time on every call, annoy genuine customers and are weak against fraudsters who buy the answers from data breaches. Callers who fail the questions create extra handling time, and callers who pass them by guessing create fraud losses.

Voice biometrics moves the check into the conversation itself. The customer enrols a voiceprint once, with consent. On later calls the system compares the caller's voice with the stored voiceprint while the customer talks about their request, and the agent sees a match or a warning instead of asking questions. Some operators also keep a library of voiceprints of known fraudsters and check new calls against it. Cloned voices are the main risk: in 2023 a journalist used an AI clone of his own voice to get into a government self service account, which is why operators add checks for live and synthetic voices and keep another factor for high value actions.

How does it work?

  1. Enrol with consent. The customer agrees to a voiceprint and records a short passphrase or free speech, in a call or an app. The voiceprint is stored as biometric data under its own retention rule.
  2. Compare on the call. In the first seconds of a later call the system extracts a voiceprint from the caller's speech and scores it against the enrolled one, with the customer's phone number and call data as extra signals.
  3. Check the voice itself. A second model estimates whether the voice is live, a replay or synthetic, and a third step can compare the voice with a library of fraudsters.
  4. Show the result to the agent. A match lets the agent skip the security questions. A low score or a flag sends the call to the normal questions or to a fraud team. The agent or the process, not the score alone, decides on any high value request.
Audience
Customer facing
Autonomy
Assist
Adoption
Mainstream
Channels
Phone and voice

What is it worth?

Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.

Value benchmarks for AI voice biometric caller authentication with synthetic voice detection
KPIMedianReported rangeData pointsClaimed by
Users servedNot pooled
1.2 million to 4.2 million
32 organization, 1 independent
Fraud losses preventedNot pooled
about GBP 249
11 organization

Value drivers: Risk and loss reduction, Speed and cycle time, Lower cost to serve, Customer experience.

Indicative value

A retail bank with 2 million inbound service calls a year

USD 66,667 to USD 750,000

Agent time released from security questions per year

How this is calculated

Formula: calls * enrolledShare * secondsSaved / 60 * costPerMinute. The low scenario uses every low input, the high scenario every high input.

InputLowHighBasis
Inbound calls that need identity verification calls, calls per year2,000,0002,000,000The reference bank.
Share of those calls from customers with a voiceprint enrolledShare, fraction of calls0.20.5Editorial assumption, replace with your own enrolment rate. HSBC UK reports more than 2.8 million active Voice ID customers, a figure that comes from a much larger customer base.
Seconds of security questions skipped per verified call secondsSaved, seconds per call2045Editorial assumption, replace with your own measured time spent on knowledge based questions.
Fully loaded cost of an agent minute costPerMinute, USD per minute0.51Editorial assumption, replace with your own contact center cost.

What it leaves out: Handling time only. It leaves out the cost of the biometric license and enrolment campaign, the fraud losses avoided or caused by false accepts, and the cost of calls that fall back to questions.

Which organizations already do this?

3 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.

This library is not a list of Blits.ai customers. It documents public AI deployments from any vendor, each with its source.

HSBC UK

United Kingdom · Banking · 2021

ScaledGrade B

HSBC UK uses Voice ID, a voice biometrics system, to check in telephone banking whether the voice of a caller matches the voiceprint held on file for the customer. The system also holds a library of fraudsters' voiceprints that new incoming calls are checked against. In May 2021 the bank reported that it was used by more than 2.8 million active customers, with around 14,000 more enrolling each week, and that reported telephone banking fraud had halved in a year. The figures are the bank's own and the release gives no method for the amount it says was prevented.

  • Fraud losses prevented: about GBP 249, in the last year, in millions
    "HSBC UK’s innovative voice biometrics system, Voice ID, has prevented almost £249 million of customers’ money from falling into the hands of telephone fraudsters in the last year, with the rate of attempted fraud down 50% year-on-year, the bank announced today."
    Claimed by: organization
  • Users served: at least 2.8 million, active customers in May 2021
    "As more people turn to digital channels and telephone banking, HSBC UK has seen an increase in customers signing up to Voice ID, used within its telephone banking service to make it easier, faster and more secure and currently used by over 2.8 million active customers."
    Claimed by: organization

Australian Taxation Office

Australia · Government and public sector · 2020

ScaledGrade C

The Australian Taxation Office lets taxpayers enrol a voiceprint and use it to prove who they are in its contact centre and in its mobile app, with a Nuance voice biometric service. By May 2020, according to a trade press report that does not name its source, it had enrolled more than 4.2 million voiceprints since 2014, three times the number at Services Australia. In 2023 the AI Incident Database logged a voice cloning test against the voiceprint system used by Services Australia and the ATO (incident 523). The article gives no outcome measure such as time saved or fraud stopped.

  • Users served: at least 4.2 million, voiceprints enrolled since 2014
    "Canberra’s other big service agency to use a Nuance voice biometric authentication service, the Australian Taxation Office, has enrolled over 4.2 million voiceprints since 2014 - more than three times that of Services Australia."
    Claimed by: independent

Services Australia

Australia · Government and public sector · 2020

ScaledGrade C

Services Australia lets customers enrol a voiceprint to verify who they are when they call for government services such as Centrelink. In answers to Senate estimates it reported 1.2 million enrolled voiceprints in March 2020. In March 2023 a Guardian Australia journalist used four minutes of his own voice, cloned with an AI tool, to get into his Centrelink self service account, which was logged as AI Incident Database incident 523. The agency defended the system and said it was secure.

  • Users served: 1.2 million, enrolled voiceprints as at 19 March 2020
    "“As at 19 March 2020, 1.2 million customers have enrolled a voiceprint with Services Australia,” it said in answers to questions on notice from senate estimates."
    Claimed by: organization

How do you implement it?

A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.

Data you need

  • Customer consent records and a retention rule for voiceprints
  • Call audio from the telephony platform in a format the biometric engine accepts
  • A set of genuine and fraudulent calls to measure false accept and false reject rates

Systems to integrate

  • Telephony or contact center platform that can stream audio
  • Agent desktop and customer record, to show the result
  • Fraud case management, for flagged calls and a fraudster voiceprint library

Complexity: High

The matching and synthetic voice detection come from a specialist biometric product. The build needs an enrolment flow with recorded consent, a connection to the telephony platform that streams call audio, an interface to the customer record and the fraud case system, a legal basis for processing biometric data in every country served, and a test of false accept and false reject rates on the organization's own callers.

  1. 1

    Settle the legal basis first

    Decide the consent wording, the retention period and the lawful basis for biometric data with the privacy team before the first voiceprint is stored. Check each country where callers live.

  2. 2

    Measure error rates on your own callers

    Run the engine silently on live calls and compare its scores with the outcome of the questions. Set the threshold from the measured false accept and false reject rates, not from the vendor default.

  3. 3

    Add checks for live and synthetic voices

    Test the system against recordings and cloned voices made with current tools, and repeat the test on a schedule, because the tools improve.

  4. 4

    Roll out by request risk

    Use a voice match to skip questions for low risk requests. Keep a second factor for payments, changes of address and other actions that fraudsters want.

Guardrails

  • Voice alone never authorizes a high value or irreversible request
  • Enrolment only with recorded consent, and an easy way for a customer to delete the voiceprint and use questions instead
  • A fallback to the standard questions for every caller who fails or declines
  • Regular tests with cloned and replayed voices, with the result reported to the fraud and privacy owners

KPIs to instrument

  • False accept rate and false reject rate measured on labelled calls
  • Share of callers who are enrolled and share of verified calls that skip questions
  • Seconds of authentication time per call, before and after
  • Fraud losses on telephone channels and the number of impostor calls caught

Human in the loop

The agent sees the match result and decides how to proceed. A flagged call goes to a fraud analyst, and a person owns the threshold, the enrolment policy and the response to a successful clone attack.

Common failure modes

Cloned voice passes the check
A generated voice matches the enrolled voiceprint, as the 2023 test on a government service showed. Add live and synthetic voice detection, keep another factor for sensitive requests and test with new cloning tools every quarter.
Threshold too tight
Genuine customers with a cold, a noisy line or a different phone are rejected and fall back to questions, which removes the saving. Review rejection by device and line type and tune the threshold.

What are the risks and rules?

EU AI Act

Depends on design

Annex III point 1(a) makes remote biometric identification systems high risk but excludes biometric verification whose sole purpose is to confirm that a person is who they claim to be. A one to one check of a caller against the voiceprint the same customer enrolled is verification, so it is outside that point. Matching callers against a library of fraudsters is identification, which can count as remote biometric identification and then is high risk, and a national law may also restrict it. Voice biometrics is in every case biometric data, which needs a condition under GDPR Article 9.

Rules that apply

Guidance

Controls to put in place

  • Data protection impact assessment before launch
  • Recorded consent per voiceprint and a documented route to withdraw it
  • Audit trail of every match score, decision and fallback
  • Scheduled tests against replayed and synthetic voices

When it went wrong elsewhere

Frequently asked questions

How many people use voice authentication in practice?

HSBC UK reported more than 2.8 million active Voice ID customers in telephone banking in May 2021. The Australian Taxation Office is reported to have enrolled more than 4.2 million voiceprints since 2014, and Services Australia 1.2 million by March 2020. Enrolment depends on customers agreeing to it.

Can an AI cloned voice fool voice authentication?

It has happened. In 2023 a Guardian Australia journalist used a clone of his own voice to enter his Centrelink account, an incident logged as number 523 in the AI Incident Database. Nuance, the supplier to both Australian agencies, says its technology can detect whether a voice is recorded or live. An organization should keep a second factor for sensitive requests and test its own system against cloned voices on a schedule.

Is voice biometrics high risk under the EU AI Act?

It depends on the design. Confirming that a caller is the enrolled customer is biometric verification, which Annex III point 1(a) excludes from the high risk category. Identifying callers against a library of fraudsters can count as remote biometric identification, which is high risk. In both cases the voiceprint is special category data under GDPR Article 9.

How to cite this page

Blits.ai AI Use Case Library, "AI voice biometric caller authentication with synthetic voice detection", last verified 18 November 2026, https://www.blits.ai/ai-use-cases/caller-voice-biometric-authentication. Licensed under CC BY 4.0. Method: how we verify use cases.

Changelog
  • 18 November 2026: First published

Related use cases

BankingPayments and cards

AI for application and identity fraud detection

Application and identity fraud detection is AI that checks incoming account and loan applications for banks and lenders to catch forgery and fraud rings. It analyzes documents, device and application data across the whole queue and cross checks against bureau and official sources to flag forged or AI generated documents, synthetic and stolen identities, and coordinated rings.

Deployments
10 public, best grade A
Reported detection improvement
2.5x
Department for Work and Pensions, organization claim
Cross industryBanking

AI quality and compliance monitoring of every customer interaction

Call quality and compliance monitoring is automated quality assurance that transcribes and scores every customer interaction, voice and chat, against the organization's own rubric. It checks required disclosures and script adherence, flags conduct and mis selling risk, and surfaces coaching opportunities, instead of the small sample a human QA team can review.

Deployments
14 public, best grade B
Reported quality score uplift
about 10%
British Gas, vendor claim
Telecommunications

AI spam and scam call blocking for mobile and landline subscribers

Spam and scam call blocking is AI in the operator's network that protects subscribers from unwanted calls by analyzing incoming calls in real time. It blocks known fraudulent calls and labels suspected scam, spam and spoofed calls on the customer's screen before they answer, so subscribers can decide whether to pick up. Fraud against the operator itself, such as SIM swap or revenue share fraud, is a separate use case.

Deployments
6 public, best grade B
Autonomy
Autonomous
Cross industryBanking

AI agent for first line contact center service

An AI first line contact center agent answers inbound customer contact on phone, chat and messaging in any sector, resolving general questions (products, fees, procedures, the status of a request) in the customer's own language and routing each caller to the right team with the context attached. Transactions on the customer's own account or contract in one sector, such as blocking a card, changing a policy or taking a meter reading, belong to that sector's servicing agent, and anything complex, sensitive or regulated goes to a person.

Deployments
116 public, best grade A
Median containment rate
52%
29 deployments, 2 "up to" values left out