ElevenLabs turned its dubbing model into a live phone translator, Deutsche Telekom is already using it

What ElevenLabs shipped, and when

Four dated events, all from ElevenLabs' own blog and public API changelog, trace a single move: the model behind its dubbing business got faster, then got wrapped in a new product aimed at live calls instead of edited video.

March 2026, Mobile World Congress. ElevenLabs and Deutsche Telekom unveiled the Magenta AI Call Assistant, described by both companies as network embedded rather than app based: it works on any phone, including landlines, because the AI runs inside the carrier's network. A caller activates it by saying "Hey Magenta." Every participant on the call is notified the moment it engages, and Deutsche Telekom says customers must opt in before it can be used at all. Telekom's own newsroom release lists live translation, call summaries and contextual support as the three features, with rollout to German customers planned for later in 2026 and support for up to 50 languages planned over the following 12 months.

28 September 2026. ElevenLabs introduced Eleven v4 and Eleven v4 Turbo across ElevenAgents, ElevenCreative and the API. v4 is pitched as the higher quality, more expressive model; v4 Turbo trades some of that for speed, with a stated median inference latency of about 100 milliseconds. Both cover more than 90 languages. ElevenLabs' own footnotes say the model was ranked No. 1 on Artificial Analysis's Provider Voice Arena Leaderboard that month, and that in blind head to head listening tests, also run that month, graders preferred Eleven v4 over four named competitors, Cartesia Sonic 3.6, Inworld TTS-2, Google Gemini 3.8 Flash-Lite TTS and Google Gemini 3.8 Flash TTS, about 75 percent of the time, with ties scored as half a point for each side.

19 October 2026. ElevenLabs' public API changelog documents the technical shape of a Realtime Translation WebSocket endpoint: a caller supplies target_language and voice_id; source_language is optional because the input language is auto detected; an eagerness parameter lets an integrator trade earlier, rougher translation against waiting for more context; enable_zrm turns on Zero Retention Mode for sensitive calls; the default text to speech model is v4 Turbo, swappable to eleven_v3 or eleven_flash_v2_5.

21 October 2026. ElevenLabs' blog publicly announces Eleven Realtime Translation: more than 90 languages, more than 17,500 voices in the library to choose from, and availability restricted to ElevenAPI enterprise customers during this alpha, who are told to contact sales rather than sign up directly. The post names Deutsche Telekom specifically as already using it to build new calling features on top of the Magenta AI Call Assistant, and quotes Deutsche Telekom VP Consumer AI Lena Drubel saying it is about "how voice can actually work and connect people, and also break down barriers."

The dated figures, lined up

Every row below traces to a primary source: ElevenLabs' blog, ElevenLabs' public API changelog, or Deutsche Telekom's newsroom. Where a number was reported only by a third party, or not reported at all, the table says so rather than guessing.

The 75 percent figure is a vendor test against four models ElevenLabs itself chose to name. That does not make it false, the methodology (blind pairings, ties as half a point) is more specific than most vendor claims, but it is not an independently reproduced benchmark, and we found no third party rerun of it. Treat it as a marketing number with a disclosed method, not an audited score.

The gap between the two blog posts, 28 September and 21 October, is 23 days. That is the entire span between ElevenLabs shipping a faster model and turning it into a telecom grade live translation product with a named customer already using it. For a company whose video dubbing product took years to reach its current language count, that is a fast turn toward a different delivery format for the same underlying speech model.

Why a live call and a dubbed video need different machinery

AI video dubbing is a batch process: a finished video goes in, and a finished, re voiced video comes out, with no one waiting on the line. Eleven Realtime Translation is the opposite: both speakers are on the call right now, and every added hundred milliseconds is felt as a conversational lag. The same underlying speech model, v4 Turbo, gets used two different ways because the deadline is different.

This is also where realtime translation can fail in a way dubbing cannot: a batch dubbing pass can always re read the full sentence before committing to a translation, but a live call, tuned for low latency, is guessing at meaning before the speaker finishes. ElevenLabs' own documentation names this tradeoff directly, recommending lower eagerness for calls like billing or claims disputes where a wrong guess costs more than a half second of extra silence.

What this means if you buy or build AI dubbing

Your dubbing vendor's engineering attention is not guaranteed to stay on dubbing. ElevenLabs shipped a telecom grade, enterprise gated live translation product 23 days after its latest speech model, with a named launch customer in Deutsche Telekom. That is a real product bet, and it competes for the same model and infrastructure team that maintains Dubbing Studio. Buyers with a dubbing roadmap tied to ElevenLabs should ask directly whether Dubbing Studio will inherit v4 Turbo, and when, rather than assuming it already has.

Zero Retention Mode is a compliance lever worth asking about by name. enable_zrm on Realtime Translation discards call audio after the session. If your dubbing contracts involve sensitive source audio, ask whether the same retention control exists, or can be made to exist, on the dubbing side, rather than assuming a general privacy policy covers it the same way.

Alpha and enterprise only means this is not yet a buying decision. There is no public pricing, no self serve signup and no stated general availability date for Realtime Translation. The news here is directional, where the leading dubbing vendor's engineering effort is pointed, not a product you can procure today.

The 100 millisecond figure is about the speech model, not about the finished dubbed video you ship. Our judgement: a 100 ms inference latency claim is meaningful for a live phone call where a human is waiting, but it says nothing about end to end turnaround for a batch dubbed video, where a translation API call, a lip sync pass and a render queue dominate total time far more than any one model's per utterance latency. Do not let a live latency number stand in for a dubbing throughput number in a vendor comparison.

Check it yourself

Every figure above comes from a page you can load directly. ElevenLabs' changelog and blog are public, no login required for reading.

# 1. The Realtime Translation announcement itself, dated 21 Oct 2026
curl -s "https://elevenlabs.io/blog/realtime-translation" -A "Mozilla/5.0" \
  | grep -io "90+ languages\|17,500\|Deutsche Telekom\|alpha" | sort -u

# 2. The Eleven v4 / v4 Turbo launch post and its footnoted methodology
curl -s "https://elevenlabs.io/blog/eleven-v4-turbo-in-elevenagents" -A "Mozilla/5.0" \
  | grep -io "100 ms\|~100\|No\. 1\|Artificial Analysis\|75%\|Cartesia Sonic\|Inworld" | sort -u

# 3. The actual API parameters, from the public changelog
curl -s "https://elevenlabs.io/docs/changelog/2026/10/19" -A "Mozilla/5.0" \
  | grep -io "target_language\|voice_id\|eagerness\|enable_zrm\|eleven_v4_turbo\|eleven_v3\|eleven_flash_v2_5" | sort -u

# 4. Deutsche Telekom's own account of the Magenta AI Call Assistant
#    https://www.telekom.com/en/newsroom/latest-updates/media-information/2026/3/deutsche-telekom-reimagines-phone-calls-with-ai-embedded-in-the

If any of those pages have changed their numbers since this was written, that is itself informative: alpha stage products are the ones most likely to have their claims revised before general availability.

What would prove this wrong

The claim under test is that ElevenLabs is putting real, near term product weight behind live, network embedded translation, as a distinct push alongside its existing async dubbing business, rather than this being a minor alpha experiment. This is wrong if, by 1 May 2027, Eleven Realtime Translation is still alpha only, with no public self serve pricing and no general availability announcement, and Deutsche Telekom's Magenta AI Call Assistant has not expanded past Germany or past a small fraction of its originally stated 50 language roadmap. A quiet, stalled alpha a year after launch would mean this was a demo, not a strategy.

A second, narrower prediction, marked as judgment: ElevenLabs will publicly confirm, by the same date, that its Dubbing Studio or Dubbing v2 API also runs on Eleven v4 Turbo or a successor realtime capable model, rather than keeping dubbing on an older model stack while the live product gets the newest one. If ElevenLabs' own documentation still does not name the model behind Dubbing Studio by 1 May 2027, treat that silence as evidence the two product lines are not actually converging.

FAQ

What exactly did ElevenLabs launch on 21 October 2026?

Eleven Realtime Translation, a live two way voice translation product delivered over a WebSocket API. Each speaker talks in their own language and hears the other speaker in theirs, in the same call, using a single consistent voice for the session. It runs on Eleven v4 Turbo by default and covers more than 90 languages with automatic language detection. At launch it is in alpha, limited to ElevenAPI enterprise customers, who are directed to contact sales rather than self serve.

What is Eleven v4 Turbo's measured latency, and who ranked it first?

ElevenLabs' own 28 September 2026 blog post states a median inference latency of about 100 milliseconds for v4 Turbo, and says the model ranked No. 1 on Artificial Analysis's Provider Voice Arena Leaderboard that same month. In a separate blind head to head preference test, also run that month by ElevenLabs, listeners picked Eleven v4 over four named competitors, Cartesia Sonic 3.6, Inworld TTS-2, Google Gemini 3.8 Flash-Lite TTS and Google Gemini 3.8 Flash TTS, about 75 percent of the time, with ties counted as half a point. These are the vendor's own figures; we found no independently run benchmark using the same method.

Is Deutsche Telekom's Magenta AI Call Assistant the same thing as Eleven Realtime Translation?

No, but they are connected. Deutsche Telekom and ElevenLabs unveiled the Magenta AI Call Assistant at Mobile World Congress in March 2026, a network embedded assistant activated by saying Hey Magenta during a call, offering live translation, call summaries and contextual support, with German rollout planned for later in 2026 and up to 50 languages planned over the following 12 months. Eleven Realtime Translation, launched in October 2026, is the underlying translation capability opened up as a standalone enterprise API product. ElevenLabs' October blog post names Deutsche Telekom as already using it to build new calling features.

Does this affect ElevenLabs' existing AI video dubbing product?

Not directly, and we could not verify whether Dubbing Studio or the Dubbing v2 API run on the v4 Turbo model at all; ElevenLabs has not said so publicly. What changed is where the company is putting a working, fast speech model: into a live, audio only, phone call product gated to enterprise customers, rather than into its asynchronous per video dubbing workflow. For buyers of AI dubbing, the relevant fact is a roadmap signal, not a feature change to the dubbing product itself.

Sources

1. ElevenLabs, Realtime Translation, blog post dated 21 October 2026. 90+ languages, 17,500+ voices, alpha for ElevenAPI enterprise customers, Deutsche Telekom named as live user, quote from Lena Drubel.
2. ElevenLabs, API changelog, 19 October 2026. Realtime Translation parameters: target_language, voice_id, source_language auto detect, eagerness, enable_zrm, default model and switchable alternatives.
3. ElevenLabs, Introducing Eleven v4, blog post dated 28 September 2026. ~100 ms median inference latency, 90+ languages, No. 1 Artificial Analysis Provider Voice Arena Leaderboard ranking, 75% blind preference test against Cartesia Sonic 3.6, Inworld TTS-2, Google Gemini 3.8 Flash-Lite TTS and Google Gemini 3.8 Flash TTS.
4. ElevenLabs, ElevenLabs and Deutsche Telekom unveil the world's first network integrated AI call assistant, blog post referencing the March 2026 Mobile World Congress unveiling.
5. Deutsche Telekom, Deutsche Telekom reimagines phone calls with AI embedded in the network, newsroom release, March 2026. "Hey Magenta" activation, opt in requirement, live translation / call summaries / contextual support, German rollout and 50 language roadmap.

Related BLOMEGA guides: Four courts, four different answers on AI voice cloning consent in 2026 · A dubbed minute costs $0.33 to $9.00, and the watermark discount is gone.