ElevenLabs opens live voice translation
On 21 October 2026 ElevenLabs launched Eleven Realtime Translation, a model that translates a live conversation: each person speaks their own language and hears the other in theirs, in a voice chosen from the company's library or cloned. It runs on Eleven v4 Turbo in more than 90 languages and is in alpha for enterprise customers of the ElevenLabs API; Deutsche Telekom is named as an early user.
Why it matters
Live interpretation, so far a human service or a text-first pipeline, becomes an API product from a speech-model vendor, aimed at telephone networks, contact centres and live events. The post names the uses and one operator but gives no latency figure, no quality measure and no price, and the alpha is open only to enterprise customers who contact sales, so what a caller would hear cannot yet be judged from outside.
What the post says. Blog post of 21 October 2026 by Eric Michaelis of the ElevenLabs B2B marketing team. The model translates while the conversation is going on; a deployment can be tuned between higher-quality translation, for complex calls such as billing or claims, and the lowest latency. It uses Eleven v4 Turbo, described as the company's low-latency speech model, "in more than 90 languages with automatic language detection". The voice is chosen from "more than 17,500 voices in our library" or built as an Instant or Professional Voice Clone, and "will stay consistent for the entire session". Deutsche Telekom is "already using" it to build calling features for its Magenta AI Call Assistant; a quotation from its VP Consumer AI is on the page. What the record does not claim. The post does not say that the speaker's own timbre, intonation or emotion is carried across; it speaks of a "natural, expressive voice" that is a chosen or cloned one. It gives no latency or accuracy number, no list of language pairs beyond the count, and no independent test. Every statement is ElevenLabs' own, and access is an alpha for enterprise API customers.