ElevenLabs launches Eleven Realtime voice agent model in 90 languages

ElevenLabs has launched Eleven Realtime, a speech-to-speech model that lets developers build a voice agent by choosing one model instead of stitching together speech recognition, conversation control, a language model and voice synthesis. It comes in two versions, Eleven Realtime v1 Mini and v1 Max, both supporting more than 90 languages, and is available through the company’s Eleven Agent platform and a real-time API.

Korean outlet Financial News reported the launch on November 3, citing industry sources. An earlier analysis by Telli dates ElevenLabs’ introduction of v1 Mini and v1 Max to October 20, describing it as an official announcement. The two dates may reflect a staged rollout, including the regional push Financial News describes for Korea, but the sources do not explain the gap.

What Eleven Realtime replaces in a voice agent stack

Until now, building a natural-sounding voice agent meant selecting and combining several separate technologies: speech recognition, voice activity detection (deciding when a person is speaking), conversational turn-taking, a large language model (LLM) and speech synthesis. Eleven Realtime integrates these into a single system, so a builder can create a voice agent simply by selecting a model. Financial News frames the aim as easing the burden on businesses and developers that build voice AI services.

“Speech-to-speech” here means the model takes spoken input and returns spoken output as one step in a live conversation. ElevenLabs has used the term before: its blog announced a public Speech-to-Speech API in March 2024. That short post does not describe conversational agents, and nothing in the sources links it to Eleven Realtime.

ElevenLabs’ agent documentation shows what the change means in practice. Language settings still apply to agents using a Realtime model, but the agent’s separate text-to-speech model setting does not apply in any language. Telli describes the split in more detail. In Custom mode, a builder picks the language model, speech model and recognition settings. Realtime mode hands those choices to ElevenLabs while keeping the agent’s instructions, voice, tools and other supported conversation settings.

Mini versus Max: two models for two kinds of work

The two versions split along speed and task complexity:

• Eleven Realtime v1 Mini focuses on fast, natural conversation. Financial News lists AI assistants, educational tutors, and characters for games and entertainment as suitable uses. Telli describes Mini as targeting everyday conversation at lower cost.
• Eleven Realtime v1 Max is specialised for complex tasks. Its listed uses are multistep work such as customer support, lead qualification and retrieving information with external tools. Telli says Max targets complex instructions and multi-step work.

Users can switch between the two by changing the model selection to suit the service. Both versions support natural turn-taking, responses when a user interrupts, and external tool calls, meaning the agent can call outside systems such as a booking or lookup service during a conversation.

Spec	Eleven Realtime v1 Mini	Eleven Realtime v1 Max
Focus	Fast, natural conversation	Complex, multistep tasks
Languages	More than 90	More than 90
Voice library	More than 17,000 voices	More than 17,000 voices
Interruptions and tool calls	Supported	Supported
Big Bench Audio (Artificial Analysis)	Not reported	98.8%
Full Duplex Bench (Artificial Analysis)	95.8%	94.1%
Time to begin answering	Not reported	0.57 seconds

Benchmarks: what the reported scores show

Financial News reports results from Artificial Analysis, a third-party evaluator, as of October:

• Big Bench Audio, which measures reasoning through voice: Max scored 98.8%.
• Response delay: Max took 0.57 seconds to begin answering after a question ended.
• Full Duplex Bench, which measures how a model handles silence, turn-taking and interruptions in conversation: Mini scored 95.8% and Max 94.1%.

Three caveats apply. First, the article’s performance chart is credited “Courtesy of ElevenLabs,” so the figures reached the press through the company, and the sources reviewed here do not include Artificial Analysis’s own published results. Second, the report gives no scores for competing voice models, so there is no way to judge from these sources whether the numbers lead the field or match it. Third, the slight edge Mini holds over Max on Full Duplex Bench fits its positioning as the conversational model. The report does not explain why the smaller model scores higher, or give a Mini score for Big Bench Audio or latency.

Telli makes a broader point about what such benchmarks leave out. A voice can sound excellent while the exchange still fails, for example when an assistant starts talking during a user’s pause or finishes a sentence it had prepared after the user corrects a name. Its example is a booking assistant that pronounces a customer’s name perfectly but reserves the wrong day because it missed a correction. Turn-taking scores measure the first problem. Task completion, which the reported benchmarks do not cover, measures the second.

Access: where Eleven Realtime is available

The models are offered through Eleven Agent, ElevenLabs’ voice agent platform, and through a real-time API. Telli places Realtime inside a wider company structure of three platforms: ElevenCreative for content creation, ElevenAgents for deployed conversational agents, and ElevenAPI for developers.

Availability is not yet universal. According to Telli, access is rolling out gradually, so the Realtime setting may not appear in every workspace yet, and existing agents must meet compatibility and regional constraints before switching. The sources do not list which regions are covered or what the constraints are. None of the sources gives pricing for Eleven Realtime, beyond Telli’s description of Mini as the lower-cost option.

ElevenLabs is also aiming the launch at specific markets. Financial News reports that the company plans to support Korean businesses and developers using voice AI, and expects it to target the enterprise voice agent market with multilingual support and simpler development. A company representative told the paper:

“Through Eleven Realtime, which supports more than 90 languages, including Korean, we plan to help Korean businesses and developers build voice agents that can converse naturally with customers.”

What it means for builders

The main change is fewer moving parts to tune. Telli argues that the most promising part of the release is less time spent adjusting the joins between components, and that a managed setup gives a small team a coherent starting point and fewer combinations to troubleshoot. A team can spend more of its effort on the questions customers ask, the information the business actually holds, and when a human should take over.

The trade-off is control. Telli notes that fewer settings mean fewer independent substitutions. If one behaviour is unsuitable, a team needs to know whether it can adjust it, change models, or return to a custom configuration. It contrasts this with OpenAI’s voice-agent guidance, which separates realtime speech interaction from a chained workflow that transcribes speech, runs a text agent and then generates speech. The chained approach lets teams inspect the intermediate text and swap individual components. Telli stresses that this is an architectural comparison, not evidence about Eleven Realtime’s internals, which have not been disclosed. For services that need a reviewable text checkpoint, such as specialist support desks or training services, that difference matters.

Some platform behaviour carries over. ElevenLabs’ agent language guide says the language for a call is fixed for its duration, so users cannot switch languages mid-conversation, and an optional language detection tool can switch to a user’s preferred language. The guide does not say whether Realtime models change this behaviour, so builders planning multilingual calls should test it directly.

For context, ElevenLabs was founded in 2022 by Mati Staniszewski and Piotr Dąbkowski and is known for expressive, multilingual voice and audio models. Its existing catalogue on third-party platforms includes text-to-speech, music generation, voice design and cloning, and Scribe speech-to-text across 99 languages. Eleven Realtime moves the company from supplying voice components to supplying the whole conversational loop.

Open questions and what to watch next

• Independent comparisons. The reported Artificial Analysis scores have no rival figures alongside them. Watch for leaderboard placements against other speech-to-speech systems.
• Task outcomes, not just turn-taking. Telli argues the real test is a messy, ordinary customer request that survives corrections and ends with the right result. Hands-on reports from deployments will be the evidence to look for.
• Rollout scope. Which workspaces and regions get Realtime mode, and what compatibility constraints existing agents face, remain unspecified.
• Configurability. Whether teams can adjust individual behaviours inside Realtime mode, or must fall back to Custom mode, will shape how widely it replaces assembled stacks.
• Pricing and language quality. No source gives Eleven Realtime prices, and none breaks down quality across the 90-plus languages beyond naming Korean and Japanese.