ElevenLabs vs other voice AI vendors: DPA comparison for compliance buyers

TL;DR. ElevenLabs and the cloud voice products (OpenAI, Azure Speech, Google Cloud Text-to-Speech) all require some form of consent from the person whose voice is cloned. The cloud providers limit custom voices to approved or allow-listed customers, while ElevenLabs offers cloning on its self-serve plans with a voice verification step. Three buyer-side controls apply whichever vendor is chosen: documented consent from any person whose voice is cloned, disclosure of deep fakes under EU AI Act Article 50(4), and tier discipline. At ElevenLabs, only the Enterprise tier offers no training on customer data by default, a choice of storage location, Zero Retention Mode and a BAA, although the DPA covers the self-serve plans too. Azure Speech and Google Cloud TTS are covered by Microsoft's and Google's existing cloud agreements, including their BAA terms. OpenAI's own voice products are contracted with OpenAI, and the Azure OpenAI route is contracted with Microsoft.

The voice AI compliance frame

Voice AI splits into three buyer scenarios with materially different compliance pictures:

1. Synthetic voice generation from text (text-to-speech with stock voices). Lower compliance load. The provider must mark the output as artificially generated in a machine-readable format (EU AI Act Article 50(2)). The deployer's deep-fake disclosure duty under Article 50(4) applies only where the output resembles an existing person (or object, place, entity or event) and would falsely appear authentic.
2. Voice cloning from a sample of a real person's voice. High compliance load. The processing needs a lawful basis under GDPR Article 6. Where the voice data is processed to identify the speaker uniquely, it is special-category biometric data (Articles 4(14) and 9(1)) and also needs an Article 9(2) condition, such as the person's explicit consent (Article 9(2)(a)). Cloning also brings defamation and IP exposure, and the EU AI Act deep-fake disclosure duty where the output is presented as that person speaking.
3. Speech-to-text transcription. Different vendor category (Whisper, Azure Speech-to-Text, Google Speech-to-Text). Compliance load is the standard processor analysis under the DPA.

This comparison addresses scenarios 1 and 2. The full ElevenLabs read is at the ElevenLabs profile.

1. The consent regime for voice cloning

| | ElevenLabs | Hyperscaler voice (OpenAI / Azure / Google) | |---|---|---| | Voice-clone product available | Yes, Instant Voice Cloning, Professional Voice Cloning | OpenAI custom voices (limited to eligible customers); Azure Custom Neural Voice (gated); Google Chirp 3: Instant Custom Voice (allow-listed users only) | | Consent verification process | Instant and Professional cloning both include a voice-captcha verification step. ElevenLabs says it confirms that the requester was present, but cannot guarantee that the recording belongs to them. Instant cloning also asks the user to confirm that they have the right and consent to clone the voice. Professional cloning works only on the account holder's own voice | The voice owner records a consent statement, which is matched to the training sample (Microsoft uses speaker verification for the match). Access is limited to approved customers | | Vendor enforcement against unauthorised cloning | Audio watermarking, checked by an Audio Detector available to signed-in users on every plan (the older classifier is now only a fallback). A "no-go voices" block on clones of prominent public figures. An Unauthorized Use of Voice report. Enforcement action against accounts, from warnings to bans | Restricted release (Microsoft-managed customers, Google allow-list, OpenAI eligible customers), with contractual restrictions on use. Microsoft watermarks personal-voice output | | Documented buyer obligation | A warranty that the buyer's content and voice models will not violate anyone's rights, plus a Prohibited Use Policy ban on cloning a living person's voice without explicit and informed consent or other valid authorisation | Microsoft: a warranty that the voice talent gave explicit written permission. Google and OpenAI: a recorded consent statement from the voice owner |

All four vendors require the buyer to obtain the voice owner's consent, in different forms. Microsoft's limited access terms require a warranty of explicit written permission from the voice talent, and Google and OpenAI require a recorded consent statement from the voice owner. ElevenLabs relies on a general rights warranty, its Prohibited Use Policy and the verification step, and it limits Professional Voice Cloning to the account holder's own voice.

The remaining compliance risk is the buyer's. A buyer who clones a voice without documented consent breaches the vendor's terms. It is also exposed under data protection law, defamation law and, where they apply, personality or publicity rights.

2. Deepfake labelling under the EU AI Act

EU AI Act Article 50(4) applies to deployers of AI systems that generate or manipulate audio constituting a deepfake. The obligation is to disclose that the content is artificially generated or manipulated. A deep fake is content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful (Article 3(60)).

Article 50 has applied since 2 August 2026, the Act's general date of application (Article 113; a fixed date). Separately from the deployer's duty, Article 50(2) requires the provider to mark synthetic audio in a machine-readable format so that it is detectable as artificially generated. Providers whose systems were on the market before 2 August 2026 have until 2 December 2026 to meet Article 50(2) (Article 111(4), inserted by Regulation (EU) 2026/1744; a fixed date).

| | ElevenLabs | Hyperscaler voice | |---|---|---| | Provider marking duty (Article 50(2)) | Yes, from 2 August 2026. For systems already on the market, by 2 December 2026 (Article 111(4)) | Yes, same | | Underlying technology subject to Article 50(4) | Yes, voice cloning produces deepfake-class output | Yes, custom voice products produce deepfake-class output | | Vendor-provided watermarking / provenance signal | Audio watermarking, checked by the Audio Detector (with the older classifier as a fallback) | Microsoft watermarks personal-voice audio and applies C2PA to avatar video. Google Cloud TTS and OpenAI: not confirmed in this review (checked 15 October 2026) | | Buyer-side disclosure obligation | Yes, on the deployer, wherever the audio is a deep fake, no later than first exposure (Article 50(5)) | Yes, same | | Exceptions | Use authorised by law to detect, prevent, investigate or prosecute criminal offences. For evidently artistic, creative, satirical or fictional works, the duty is reduced to disclosing that generated content exists, in a way that does not hamper the work | Same |

The deployer's disclosure duty is the same whichever vendor is used. A UK agency whose campaign runs in the EU owes it whether the voiceover comes from ElevenLabs or from Azure Custom Neural Voice, provided the voiceover resembles a real person and would falsely appear authentic. The AI Act reaches deployers outside the EU where the output is used in the Union (Article 2(1)(c)).

3. Training defaults and customer-data use

| | ElevenLabs Enterprise | OpenAI Voice (API) | Azure Speech Custom Neural Voice | Google Cloud TTS Custom Voice | |---|---|---|---|---| | Vendor uses cloned voices to train shared models | No (Enterprise tier) | No (unless the customer opts in) | No (tenant-scoped) | No (tenant-scoped) | | Vendor uses prompts / inputs to improve service | No (Enterprise tier) | No (unless the customer opts in) | No | No | | Customer voice clone reusable by other customers | No, unless the owner publishes a Professional Voice Clone to the Voice Library or shares it | Not confirmed in this review | No (the customer holds exclusive usage rights to the model) | Not confirmed in this review |

At ElevenLabs only the Enterprise default is protective. Data from the self-serve plans, including API use, is used for training unless the customer opts out, and the opt-out applies only to data submitted afterwards. OpenAI does not train on API data unless the customer opts in. Microsoft uses a customer's custom voice training data only for that customer's models, and Google does not use Customer Data to train models without the customer's permission or instruction.

4. BAA position for US healthcare voice work

| | ElevenLabs | OpenAI Voice | Azure Speech | Google Cloud TTS | |---|---|---|---|---| | BAA available | Yes, Enterprise only, arranged through the account team. It covers the Text to Speech API, the Speech to Text API and ElevenAgents, each only with Zero Retention Mode enabled. ElevenAgents configuration data is outside the BAA | Yes. API customers can sign the BAA without an enterprise agreement once they have an established history of API usage. API HIPAA eligibility also requires Modified Retention unless OpenAI specifies otherwise, and the eligible endpoints include /v1/audio/speech, /v1/audio/transcriptions and /v1/realtime. ChatGPT Enterprise and Edu BAAs go through sales, and there is no BAA for ChatGPT Business. The Azure OpenAI route is contracted with Microsoft | Yes. Microsoft includes the BAA by default in its Product Terms for customers that are covered entities or business associates, for in-scope Azure services. Microsoft's Azure Compliance Offerings document (September 2026) lists Azure Speech in Microsoft Foundry in scope for the HIPAA BAA | Yes, Google Cloud BAA covers eligible services (Text-to-Speech is listed) | | Consumer tier eligibility | No. There is no BAA on any self-serve plan (Free, Starter, Creator, Pro, Scale or Business) | No | N/A (no consumer tier) | N/A (no consumer tier) |

For US healthcare voice work (clinical narration, patient-facing voice assistants, audio-summarisation of clinical notes), the BAA gate is the procurement floor. Consumer or creator tiers of any voice AI vendor are not BAA-eligible. See the HIPAA hub for the full checklist.

5. Subprocessor depth

| | ElevenLabs | Hyperscaler voice | |---|---|---| | Cloud subprocessors | ElevenLabs lists 32 subprocessors (15 October 2026), including Google Cloud, AWS, Azure (for LLM services where India data residency is selected) and MongoDB | Azure Speech and Google Cloud TTS run on Microsoft's and Google's own clouds | | Foundation model provenance | ElevenLabs' own voice models. ElevenAgents can also use third-party LLMs (Google Gemini, Anthropic Claude and OpenAI GPT models among them), and the subprocessor list also names xAI, Fireworks AI and Databricks for LLM services | OpenAI / Microsoft / Google first-party | | Transfer mechanism to the US entity | Eleven Labs Inc. is Active under the EU-U.S. DPF, the UK Extension and the Swiss-U.S. DPF to 9 February 2027. The DPA also provides for the SCCs | Microsoft Corporation and Google LLC are Active under all three. OpenAI is not on the DPF register, and its DPA uses the SCCs |

ElevenLabs' subprocessor list is longer because it includes third-party model providers for its LLM, image and video features. Data residency is an Enterprise option that ElevenLabs offers at its discretion through its own isolated environments. Transfers to the US entities of ElevenLabs, Microsoft and Google can rely on the EU adequacy decision for the DPF (and, for UK transfers, the UK Extension), which does not require a transfer impact assessment. Transfers to OpenAI rely on the SCCs.

6. EU/UK transfer mechanism

All four vendors publish the 2021 SCCs (Module 2 where the customer is the controller) with the UK addendum. ElevenLabs, Microsoft and Google, but not OpenAI, are also Active under the EU-U.S. DPF, its UK Extension and the Swiss-U.S. DPF (registry checked 15 October 2026). Region commitments vary:

• ElevenLabs: Enterprise data residency in the EU, India or Singapore (the US is the standard location). Residency controls where data is stored, not where it is processed. With EU residency, processing can be kept in the EU by using Zero Retention Mode through the API, unless optional integrations such as custom LLMs or post-call webhooks need processing outside the region
• OpenAI: eligible API customers and new ChatGPT Enterprise and Edu customers can choose where in-scope customer content is stored at rest; for ChatGPT the supported regions include Europe (EEA and Switzerland) and the United Kingdom
• Azure Speech: EU region available; EU Data Boundary applies on in-scope services
• Google Cloud TTS: EU multi-region and European regional endpoints (London, Frankfurt and Eemshaven). With a regional endpoint, data at rest and in use stays within Europe

Picking between them

Pick ElevenLabs Enterprise when:

• The use case is creator workflow, brand voice production, audiobook narration, podcast post-production, or accessibility narration
• The voice owner can create and verify a Professional Voice Clone on their own account and share it with the buyer, which shows that the clone came from the owner's account (it does not replace the buyer's own written record of consent)
• The buyer wants generated audio watermarked, with a detector that anyone with an ElevenLabs account can use to check it

Pick a hyperscaler voice product when:

• The buyer is already on Azure or Google Cloud and the procurement gate values single-vendor data-flow simplicity
• The use case is enterprise voice (IVR, customer support, internal voice assistants) where the integration into existing Microsoft or Google infrastructure outweighs voice quality differences
• HIPAA BAA, EU Data Boundary, or sovereign-cloud commitments are part of the procurement gate and the hyperscaler's broader framework simplifies the audit

The deployer's deep-fake disclosure duty under Article 50(4) is the same whichever vendor is chosen.

The voice-clone-without-consent failure mode

Cloning a real person's voice without documented consent is a serious compliance failure, and choosing a different vendor does not prevent it.

The downstream consequences include GDPR exposure (no lawful basis under Article 6 and, where the voice is processed to identify the speaker, no Article 9 condition), defamation and right-of-publicity claims, breach of the vendor's own terms (which the buyer warranted compliance with), and reputational damage that survives any later contractual remediation.

The buyer-side control is procedural: any voice-clone deployment has to include a documented consent artifact from the natural person whose voice is cloned, refreshed when the use case changes, and revocable on request. The vendor cannot make this happen for the buyer.

Related reading

• The DPA hub: DPA for AI vendors
• The deployer-side AI Act read covering deepfake labelling: EU AI Act for AI buyers
• For US healthcare voice work: HIPAA for AI tools
• Cross-comparison: OpenAI vs Anthropic DPA (for the underlying foundation model picture if voice runs on top of a general-purpose model)

For ongoing AI compliance support, work with Janus DPO-as-a-Service. Browse the vendor index, other vendor comparisons, or the AI Agent Incident Register.