ElevenLabs

innFactory AI Consulting from Rosenheim, Germany advises companies in the DACH region on GDPR- and AI-Act-compliant use of voice and music AI. ElevenLabs has grown from a pure text-to-speech provider into a broad audio stack: text-to-speech with Eleven v4 and Eleven v4 Turbo, speech-to-text with Scribe v2, music generation with Music v2.5, Dubbing v2 for localisation and the agent platform ElevenLabs Agents. The predecessors Eleven v3 and Eleven v3 Conversational remain available.

Practically relevant for many companies: ElevenLabs is the voice layer behind numerous European voice-agent and telephone-assistant products. Anyone using an AI phone assistant therefore often uses ElevenLabs indirectly via that vendor – the data-protection chain must then be documented down to the sub-processor, including processing region and retention period. This overview reflects the state as of October 2026.

Current models at a glance

Eleven v4 and Eleven v4 Turbo

On 28 September 2026, ElevenLabs released Eleven v4 and Eleven v4 Turbo. According to ElevenLabs, Eleven v4 improves output quality, voice fidelity, emotion, audio tags and language coverage over Eleven v3; the docs explicitly recommend switching to v4 and testing it with your own voices. The model supports 90+ languages and up to 10,000 characters per request (about 10 minutes of audio) and is used via the Text to Dialogue API.

For voice cloning, ElevenLabs cites clear improvements for Instant and Professional Voice Clones. One deliberate change concerns cross-language generation: if the target language differs from the language of the reference voice, v4 produces fluent speech in the target language instead of carrying over the reference accent. If brand voices should keep their accent, test this before migrating. Because v4 reproduces the source voice more faithfully, it can sound noticeably different from v3, and the quality of the source recording matters more.

Eleven v4 Turbo is the real-time variant of the family, with a median inference latency of about 100 ms according to the docs. It is used via the Text to Dialogue WebSocket and targets support agents, assistants and interactive characters.

Eleven v3 and Eleven v3 Conversational

Eleven v3 (GA since March 2026) generates natural-sounding speech with a wide emotional range in over 70 languages. The docs now list v3 as the previous generation. Key features:

Audio tags: emotional and acoustic instructions can be controlled directly in the text using tags in square brackets:

[excited] We are delighted, [whispers] that you are here today.
[sighs] After a long day...
[clapping] Well done!

The model interprets tags such as [excited], [whispers], [sighs], [gunshot], [clapping] or [explosion] and adapts tone and audio accordingly. With the Text to Dialogue API, multi-speaker scenarios (podcasts, audio dramas, training videos) can be generated in a single call.

Eleven v3 Conversational (May 2026) closed the gap between studio quality and real-time latency: about 280 ms latency across 70+ languages and an Expressive Mode that uses real-time signals from Scribe v2 Realtime, including prosody, to time responses more naturally. Documented limitation: characteristics of Professional Voice Clones are not reliably preserved; for cloned brand voices the docs point to the Flash models.

Language coverage compared: v4 and v4 Turbo 90+, v3 and v3 Conversational 70+, Flash v2.5 32, Multilingual v2 29, Scribe v2 and Dubbing v2 90+.

Scribe v2, Scribe v2 Realtime and Scribe v2 Medical

Scribe v2 is the batch variant for long recordings, subtitling and audio analysis: 90+ languages, speaker diarisation for up to 32 speakers, entity detection and keyterm biasing for technical terms. Scribe v2 Realtime is the streaming variant with about 150 ms latency per the model docs; it is the default ASR path in ElevenLabs Agents and supplies the signals for Expressive Mode. Entity detection is also available in the live stream. Leaner real-time variants (scribe_v2_realtime_turbo, scribe_v2_realtime_lite) are documented as well; check the current docs. Since 28 September 2026, batch and real-time transcription also accept a natural-language edit instruction (up to 2,000 characters) for post-processing the transcript.

Scribe v2 Medical (scribe_v2_medical, since 11 September 2026) is a batch model for clinical audio such as doctor-patient conversations, dictation and intake calls. According to ElevenLabs it delivers 18% fewer transcription errors on clinical audio than Scribe v2. It is HIPAA-capable (BAA for Enterprise customers) and covered by Zero Retention Mode; according to the vendor it does not replace clinical judgement – transcripts must be reviewed by qualified staff before use.

Music v2.5 and Music v2

Music v2 generates complete songs including vocals and arrangement, with section-level inpainting and mid-track genre changes. According to the vendor it was trained exclusively on licensed data and is cleared for commercial use – relevant if AI music is to be used in advertising. Via the API (model_id: music_v2) it works with chunk-based Composition Plans; a streaming endpoint returns the composition plan and optionally word timestamps.

Music v2.5 (music_v2_5) has been available since 11/14 September 2026 in ElevenMusic, Studio, Flows and the API and is, according to the vendor, the new default model for prompt and reference generation with more natural instrumentation and vocals, especially in vocal- and acoustic-heavy genres. Composition-plan chunks support up to 6,132 characters. Music v2 remains usable in parallel.

Dubbing v2

Dubbing v2 has been available via the API since 6 August 2026. The model translates content into 90+ languages and conditions directly on the original recording, so tone, emotion and delivery are preserved. The translation logic is sync-aware, and regional variants (e.g. Castilian versus Latin American Spanish) are distinguished. The API is project-based: transcript and translation are available as editable JSON, and after a correction only the changed segments are re-rendered.

Important for EU projects: according to the docs, Dubbing is not available in the isolated data-residency environments. If you want to combine dubbing with EU data residency, clarify this with ElevenLabs beforehand.

SynthID watermark

Since 25 June 2026, ElevenLabs has embedded the SynthID watermark, in collaboration with Google DeepMind, in generated audio. It is inaudible and, according to the vendor, survives cropping, tempo changes, format conversion and metadata removal. Audio generated before June 2026 carries no watermark. A free ElevenLabs Audio Detector is available for verification; it only detects watermarks from the ElevenLabs ecosystem.

ElevenLabs Agents

ElevenLabs’ conversational-AI platform (also ElevenAgents) combines four building blocks: a fine-tuned speech-to-text model, a freely selectable LLM (including your own models), a low-latency text-to-speech layer with 5,000+ voices and a dedicated turn-taking model for timing. According to the changelog, current models from OpenAI, Anthropic, Zhipu and DeepSeek, among others, are available as agent LLMs.

Channels: telephony via SIP trunking and Twilio (including batch calls), WhatsApp, web widget and React SDK, iOS/Android/React Native, and a WebSocket API.

Operations: multi-step workflows with a visual builder, knowledge base with RAG, tool integration with webhook timeouts of up to 300 seconds, configurable guardrails, A/B tests, automated agent tests including tool mocks, transcript search and versioning via CLI and MCP.

Pricing per the pricing page: USD 0.08 per minute on all plans, USD 0.16 per minute in burst mode above the concurrency limit, text messages USD 0.003. LLM and telephony costs are billed separately – a point that is regularly underestimated when budgeting phone assistants. Included minutes and concurrent conversations increase with the plan.

Pricing

Subscription plans per the pricing page (monthly billing):

Prices per the vendor, as of October 2026. Credits are drawn from a shared pool across all products; per-plan allowances and per-model consumption change with new models and should be checked against the current pricing page before signing a contract.

EU availability and GDPR

Data residency: an Enterprise feature, not the default

ElevenLabs offers isolated environments with data residency in the EU, India and Singapore; the default is US infrastructure. The docs explicitly list data residency as an Enterprise feature – without the corresponding contract, data is processed in the US default environment regardless of the customer’s location. This is the most important point for GDPR assessments in mid-sized companies.

Further restrictions:

• Dubbing is not available in isolated environments according to the docs
• Availability of LLM models in Agents varies by region and provider
• Existing resources can only be migrated to a limited extent: Professional Voice Clones can be shared into an isolated environment, while agents must be recreated there via the API. An endpoint replicates voices across data-residency environments.

Zero Retention Mode

For particularly sensitive workflows, ElevenLabs offers a Zero Retention Mode: request and response data are deleted immediately after processing instead of being stored. This mode is also reserved for Enterprise customers. Technically it is set via the query parameter enable_logging=false, or in Agents per agent in the privacy settings. According to the docs it covers text-to-speech, speech-to-text, Text to Dialogue, Voice Changer and Agents.

Two practical points: the web UI and playground are never covered – usage must be directed to the API. And support is limited in Zero Retention Mode because no diagnostic logs exist.

EU AI Act and labelling

• Transparency obligations apply to synthetic audio content under the EU AI Act; the relevant provisions apply from August 2026
• SynthID and the Audio Detector are technical proof of origin but do not replace a notice addressed to the user – an audible or visible disclosure still has to be arranged organisationally
• Recommendation: an internal policy for labelling AI audio in customer and employee communication, especially for phone assistants

Voice cloning: consent and abuse protection

• ElevenLabs uses a Voice CAPTCHA for voice cloning: whoever clones a voice has to read out a displayed text within a few seconds. This shows that the requesting person was present during the recording – it does not show that the voice belongs to them
• The terms of use prohibit non-consensual cloning, deceptive deepfakes and circumvention of the verification mechanisms
• The documented consent of the cloned person remains a separate obligation of the company
• For brand voices: a contractual arrangement with speakers is essential, including rules on continued use after the business relationship ends

When ElevenLabs runs via a third party

Many voice-agent and phone-assistant products on the European market use ElevenLabs as their voice layer without this being visible in the front end. The actual contract partner is then the product vendor, and ElevenLabs is a sub-processor. Clarify which region the vendor calls ElevenLabs in, whether it holds an Enterprise contract with EU data residency and Zero Retention Mode itself, and how long call recordings and transcripts are retained at both levels.

Model lifecycle

• eleven_monolingual_v1 and eleven_multilingual_v1: shut down on 9 July 2026 – migrate to Multilingual v2 or newer
• scribe_v1: shut down on 9 July 2026 – migrate to Scribe v2 or Scribe v2 Realtime
• eleven_turbo_v2 and eleven_turbo_v2_5: listed as deprecated in the docs, successors are Flash v2 and Flash v2.5
• Eleven v3 and v3 Conversational: still available, listed as the previous generation in the docs; successors are v4 and v4 Turbo

Anyone using ElevenLabs via a third party should check which model it actually calls – deprecated model IDs are a common cause of silent quality changes.

Model selection by use case

Integration into enterprise workflows

• REST API with extensive streaming options
• WebSocket streams for real-time conversations and live transcription
• SDKs: Python, JavaScript/Node.js, iOS, Android, React Native, plus a CLI and an MCP server
• Agents: freely selectable LLM as backbone, knowledge base with RAG, tool calls, guardrails and automated agent tests

Our recommendation

For high-quality speech output, Eleven v4 is the obvious choice, and for phone assistants and voice agents Eleven v4 Turbo – test cloned brand voices before migrating, because sound and accent behaviour can change compared with v3. For transcription we recommend Scribe v2 (batch) or Scribe v2 Realtime (in the agent), for medical audio Scribe v2 Medical, for multilingual video localisation Dubbing v2, and for AI music usable in advertising Music v2.5 (Music v2 remains available as an alternative) – both music models benefit from the licensed training data.

For GDPR-critical applications, the route goes through the Enterprise tier: EU data residency and Zero Retention Mode are available there, but not in the standard plans. Anyone using ElevenLabs only indirectly via a voice-agent product should request these points from the vendor in writing before personal call data flows. We also recommend an internal policy for labelling AI audio – SynthID provides technical traceability, while labelling towards data subjects remains an organisational task.

Depending on the use case, we also evaluate OpenAI gpt-4o-mini-tts (well integrated with the OpenAI stack) or Cartesia Sonic (designed for low latency). Contact us for advice on the right audio model strategy.