ElevenLabs Realtime Translation: API, Languages and Enterprise Alpha

ElevenLabs opened a limited enterprise alpha of Eleven Realtime Translation on October 21, 2026. It translates live speech during calls, supports more than 90 languages and detects the source language automatically. Unlike the company's voice-agent models, this release helps people talk to other people across languages.

The official announcement makes the access boundary clear: this is an enterprise alpha, not a generally available feature for every ElevenLabs subscriber. Public API documentation nevertheless gives developers enough detail to understand the integration and its tradeoffs.

How the Realtime Translation API works

The WebSocket API accepts audio chunks and streams translated output. A session needs a target language and an output voice. If the source language is omitted, the service detects it.

Alongside translated audio, applications receive translated text and interim and final source transcripts. That creates a useful separation between the spoken experience and the text a product may display as captions. Interim text is still being revised, so an interface should not present it as a final record.

The product page explains that two-way conversations are configured at the network or application level, for example by connecting each caller to a separate endpoint instance. A single translation session should not be mistaken for an entire two-way calling product.

A developer still needs to handle audio routing, playback and what happens when both people speak. This release supplies a translation layer; the surrounding call experience remains a product integration problem.

The eagerness control exposes a real tradeoff

The API offers low, medium and high eagerness, with medium as the default. High starts translating sooner. Low waits for more of the sentence, trading additional delay for better translation quality.

That is a more useful control than a universal promise of instant translation. A sentence can change meaning as the speaker continues. Starting output earlier leaves less context available when the system decides what to say.

For a customer-support call, a small delay may be acceptable if it prevents a misleading instruction. For informal conversation, participants may prefer quicker responses even when wording is less polished. Teams should test the setting against their actual language pairs and speaking patterns, rather than choosing the lowest delay automatically.

Does it preserve the original speaker's voice?

Not exactly. ElevenLabs says Realtime Translation uses a selected voice rather than automatically carrying the original speaker's voice across languages unchanged. Customers can choose a library voice or use an Instant or Professional Voice Clone to approximate a particular speaker.

Eleven v4 Turbo is the default speech-output model. The distinction matters because voice consistency and voice identity are different product requirements. A stable synthetic voice may make a conversation easier to follow without being a faithful reproduction of the person speaking.

Access, pricing and operational limits

The documentation restricts the endpoint to enterprise workspaces. The launch materials do not publish a Realtime Translation price, concurrency allowance or numerical end-to-end latency guarantee.

The public ElevenAPI pricing page lists rates for other speech products, including Realtime v1 Mini and Max. Those prices should not be reused as a quote for Realtime Translation. Enterprise buyers need confirmation of the translation product's own commercial terms.

The API also documents recoverable segment failures separately from session-ending errors. An application should plan for missing output, not only a disconnected call. If one translated segment fails, users need a clear way to notice and repeat or clarify it.

For an initial evaluation, measure translation accuracy and conversational delay together. A natural voice can make incorrect output sound convincing, so fluent audio is not enough evidence of a successful conversation.

Live translation becomes useful when people can keep speaking and correct misunderstandings easily. The important test for this alpha is whether the whole conversation works, not whether an isolated sentence sounds good.

I’m Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders turn AI launches and implementation details into clear stories that developers and buyers can understand and share.

Book a 30-minute call.