Is Voice Recognition AI High-Risk Under the EU AI Act?
Voice biometric identification (1-to-many) is high-risk under Annex III point 1. Voice verification (1-to-1) and transcription are not. Drawing the line.
It depends entirely on what the system does with the voice. "Voice recognition" is three different things wearing one name, and the EU AI Act tiers them differently:
- Voice biometric identification (one-to-many) — matching an unknown speaker against a set of enrolled voiceprints to work out who is speaking. This is remote biometric identification, high-risk under Annex III point 1.
- Voice verification (one-to-one) — confirming that a speaker is who they claim to be, by comparing their voice to a single enrolled template (a voice login). This is excluded from Annex III point 1 and is, as a tier, minimal-risk.
- Speech transcription and speaker diarisation — turning audio into text, or labelling "Speaker A / Speaker B" without establishing identity. No biometric identification, so minimal-risk.
So the same microphone, the same model family, three tiers. The decisive question is whether the system establishes a person's identity against many candidates, confirms one claimed identity, or never identifies anyone at all. This page draws all three lines and shows where each lands.
The Three-Way Split, Drawn Precisely
The Act's biometric high-risk limb is narrow on purpose. Annex III point 1(a) covers remote biometric identification systems — and a recital clarifies that one-to-one verification/authentication confirming a claimed identity is not caught by that limb.
One-to-many identification → high-risk
A call centre that takes an inbound caller's voice and searches it against thousands of enrolled voiceprints to decide which customer this is is doing one-to-many identification. So is a system that flags a caller against a watchlist of known voices. This establishes identity from a field of candidates — the definition of remote biometric identification — and it is high-risk under Annex III point 1(a), carrying the full obligation stack and the notified-body conformity route.
One-to-one verification → minimal
A bank that lets a customer authenticate by saying a passphrase, comparing the voice to that customer's own enrolled template, is doing one-to-one verification. The system answers "is this the person they claim to be?", not "who is this?". Recital 54 puts authentication outside Annex III point 1, so as a tier it is minimal-risk. (GDPR Article 9 still applies — a voiceprint is biometric data — but the AI Act high-risk stack does not.)
Transcription and diarisation → minimal
A meeting tool that transcribes speech to text, or labels turns as "Speaker 1 / Speaker 2" without linking those labels to real identities, performs no biometric identification at all. It is minimal-risk under the AI Act. The moment diarisation is wired to named identities by matching voiceprints, re-run the test — it may have become one-to-many identification.
| System | What it answers | Tier | Basis |
|---|---|---|---|
| Voice identification (1:many) | Who is speaking? | High-risk | Annex III point 1(a) |
| Voice verification (1:1) | Are they who they claim to be? | Minimal | Excluded from Annex III point 1 (recital 54) |
| Transcription | What was said? | Minimal | No biometric identification |
| Speaker diarisation (unnamed) | Who spoke when (anonymously)? | Minimal | No biometric identification |
Why the One-to-Many / One-to-One Line Is the Whole Game
The distinction is not pedantry; it tracks the harm the legislature cared about. One-to-many identification can pick a person out of a crowd of voices, often without their cooperation, and attach a name. One-to-one verification only works when the person actively presents themselves and claims an identity — it cannot trawl. That asymmetry is why authentication is carved out and identification is not.
Practically, ask three questions of any voice system:
- Is there a reference set of more than one person? If the system searches a gallery of voiceprints, it is identification. If it compares to exactly one claimed template, it is verification.
- Does the user claim an identity first? Verification starts from a claim ("I am customer 4471") and tests it. Identification starts from "unknown".
- What is the output — a name from many, or a yes/no on one? A name pulled from a database is identification; a match/no-match on a single template is verification.
If all three point to a single, claimed identity, you are in the minimal-risk verification lane. If any points to a search over many, treat it as high-risk and run the Annex III point 1 programme.
Where Voice Recognition Can Tip Into Prohibited
Two adjacent prohibitions matter, even though plain voice identification is "only" high-risk.
Real-time remote identification in public spaces (Article 5(1)(h)). Voice is a biometric. If a system performs real-time remote biometric identification of people in publicly accessible spaces for law-enforcement purposes, it is largely prohibited, subject to narrow, authorised exceptions. A live voice-trawl of a public square for police use is the prohibited case, not the high-risk one.
Sensitive-attribute categorisation from voice (Article 5(1)(g)). A system that categorises speakers by inferring a protected attribute — sexual orientation, ethnicity, religion — from vocal characteristics is prohibited categorisation, not high-risk identification. Inferring who from voice is high-risk; inferring a sensitive trait from voice is banned.
Neither bites a typical enterprise voice-login or call-routing tool — but a vendor that bolts on "emotion from voice" or "demographic profiling from voice" can walk a deployer into a different tier without anyone noticing.
Obligations If Your Voice System Is High-Risk
Once a voice system is confirmed as one-to-many identification, it is high-risk under Annex III point 1, and Annex III point 1 is the corner of the high-risk world with the strictest conformity route.
Provider (Article 16 stack): risk management (Article 9), data governance and bias testing across accents and demographics (Article 10), technical documentation (Article 11 / Annex IV), logging (Article 12), transparency and instructions (Article 13), human oversight by design (Article 14), accuracy and robustness (Article 15). Conformity assessment runs through a notified body under the Annex VII procedure — the biometrics exception in Article 43 — unless harmonised standards covering all requirements have been applied — and as of October 2026 no harmonised standard has been cited in the Official Journal. Then register under Article 49, monitor post-market under Article 72, and report serious incidents under Article 73.
Deployer (Article 26): use within the intended purpose, assign competent human oversight, keep logs for at least six months, and provide the Article 50(3) information where the use also involves biometric categorisation. Public bodies and public-service providers add an Article 27 fundamental rights impact assessment before first use. And under Article 25, a deployer that rebrands or substantially modifies the system, or changes its intended purpose so that it becomes high-risk, or repurposes it to a new high-risk use, becomes the provider.
For a verification-only or transcription system, none of this stack applies — the duty is GDPR (for any biometric data) plus, where the tool talks to people, the Article 50(1) "you are interacting with AI" disclosure.
Worked Example: TalkBridge GmbH, a Contact-Centre Vendor
TalkBridge GmbH, an Austrian contact-centre software vendor, ~120 staff, ~€18M turnover, ships two voice features and is building a third.
- Voice login. A caller says a passphrase; the system compares it to that caller's own enrolled voiceprint. One-to-one verification — excluded from Annex III point 1, minimal-risk as a tier. GDPR Article 9 applies; the AI Act high-risk stack does not.
- Live transcription. Calls are transcribed to text for agent notes; speakers are labelled "Agent / Caller" with no identity matching. No biometric identification — minimal-risk.
- Caller identification (in development). A new feature would take an inbound voice and search it against the client's whole customer voiceprint gallery to auto-route "which customer is this?". That is one-to-many identification — high-risk under Annex III point 1(a).
For the identification feature, TalkBridge is the provider and must run the full Article 16 stack plus an Annex VII notified-body assessment before placing it on the market, and register under Article 49. Its call-centre customers become deployers under Article 26. The two shipped features stay minimal-risk. One roadmap idea — "detect caller emotion to prioritise angry callers" — is not prohibited in a contact-centre context (workplace emotion recognition is prohibited for employees, not for inferring a customer's state for service routing), but emotion recognition is itself high-risk under Annex III point 1(c), so it would join the identification feature's programme and need an Article 50(3) notice and a GDPR assessment — while a sibling idea, "infer caller demographics like ethnicity from voice", would be prohibited under Article 5(1)(g) and is dropped. Mis-tiering the verification feature as "high-risk like the rest" would have saddled the whole product line with a notified-body burden it does not need; the precise split is what keeps the programme proportionate.
Deadlines: What Binds and When
The high-risk Annex III point 1 obligations for the identification feature originally read 2 August 2026 in the Regulation. The Digital Omnibus — Regulation (EU) 2026/1744, published in the Official Journal on 24 July 2026 and in force since 27 July 2026 — moved them to 2 December 2027. Plan the high-risk identification programme against 2 December 2027.
The prohibitions that bound the adjacent cases — Article 5(1)(h) real-time public identification and Article 5(1)(g) sensitive-attribute categorisation — are already in force, since 2 February 2025. The Article 50 transparency duties (50(1) and 50(3)) have applied since 2 August 2026 for everyone, not delayed. Verification and transcription carry no high-risk deadline because they are not high-risk.
How Confir Helps
Confir's rule-based engine tiers a voice system by its function, not its label. The intake distinguishes one-to-many identification (high-risk, Annex III point 1, notified-body route) from one-to-one verification and from transcription/diarisation (minimal), and it runs the Article 5(1)(h) and Article 5(1)(g) prohibition tests first so an identification trawl or a sensitive-attribute inference surfaces as a stop finding. For a confirmed high-risk identification system it structures the Article 11 / Annex IV documentation for the Annex VII assessment and maps the Article 27 FRIA where it applies. The classification is deterministic and rule-based — same intake, same finding, the rule that fired shown in plain language.
Related guides
- Biometric identification (Annex III)
- Remote biometric identification
- Article 5 prohibited practices
- Emotion recognition at work
Manage your EU AI Act compliance in one place
Confir automates risk classification, technical documentation, and audit trails for any company. No consultants. No 6-month projects. 14-day free trial.
Start free trial →