ElevenLabs Realtime Translation

ElevenLabs Realtime Translation turns live speech into natural, expressive translated audio as people talk. Each person speaks their own language and hears the other in theirs, so two parties who do not share a common language can hold a conversation without reading captions or waiting on a separate interpretation step. The site frames the promise simply as "Talk to anyone in any language." The product is positioned as a streaming voice layer for telecom operators, contact centers and BPOs, and live events, and it can output translated speech in more than 90 languages using Eleven v4 Turbo by default. Access is provided through an API with published documentation, a single WebSocket connection and per-session configuration parameters. The problem it addresses is the friction of real-time conversation across languages. Caption-only tools require participants to read rather than listen, and stitched pipelines that chain separate speech-to-text, language model and text-to-speech stages add latency and operational complexity between the spoken word and the translated audio. ElevenLabs describes Realtime Translation as going from live speech to expressive translated audio over one WebSocket, keeping a single consistent voice throughout the session. The company also distinguishes it from Eleven Realtime Dubbing: Realtime Translation is intended for more latency-critical use cases such as two-way conversation, where both parties can speak and hear the other party in their own language. Language coverage is the foundation of the product. Output is available in 90+ languages, with the site listing German, Japanese, Hindi, Spanish, French, Arabic, Mandarin Chinese, Bengali and Portuguese among them, plus accent options such as New York, Cockney, Irish, Australian, Canadian, German and Italian, alongside a note of +93 languages and accents. Voice selection is handled separately from translation: users pick from a library of over 17,500 expressive, human-like voices, or supply their own Instant or Professional Voice Clone as the voice_id. That chosen voice stays consistent for the whole conversation. ElevenLabs states explicitly that Realtime Translation does not carry the original speaker's voice across languages exactly; instead the selected output voice approximates a specific speaker, which is why choosing a clone or a library voice matters for caller experience and brand consistency. Because conversation tolerates different amounts of delay in different contexts, the product exposes an eagerness parameter per session, set to low, medium or high. High begins translating soonest for the lowest latency; low waits for more of the sentence to arrive, producing a higher-quality translation with more delay; medium is the default. Model choice is also configurable. Eleven v4 Turbo is the default text-to-speech model for translated speech, and customers can set tts_model_id to eleven_flash_v2_5 or eleven_v3 instead. ElevenLabs notes that Eleven v4 Turbo gives one voice stronger performance across multiple languages, which supports the promise of consistent voice quality from one language to the next and lets teams match the model to the latency profile of each deployment. Beyond audio, every session streams translated text alongside the speech, plus interim and final transcripts of the source audio, which can power live captions or logging on top of the translated audio. Automatic language detection identifies the language each caller speaks without manual selection. The product is telecom-ready: it works with the audio formats networks use and is easily configurable via the API. Two-way translation is supported, and bidirectionality is configured at the network or product level — for example, by connecting each caller to a dedicated instance of the endpoint, so both sides can speak their own language and hear the other in theirs. The overall approach is streaming rather than batched. Speech enters over one WebSocket connection and translated, expressive audio comes out as the conversation continues, with the same selected voice carried through the session. Because the pipeline is delivered as a single voice layer instead of several stitched services, an operator configures behaviour through parameters rather than assembling and maintaining separate components: eagerness for the speed-versus-quality trade-off, tts_model_id for the speech model, and voice_id for the output voice, all settable per session. Published API reference documentation covers the realtime translation endpoint so teams can wire the layer into their own network or product. The stated outcome is that people understand each other in the moment. Every conversation gets the same natural voice quality, and voices stay consistent across more than 90 languages, so a listener does not hear a different-sounding voice each time the language changes. Expressive delivery is part of the promise, and the site's demo illustrates emotional tones such as reassuring, empathetic, frustrated, urgent and apologetic. For businesses, the benefit is reach: telecom operators can make conversations seamless regardless of what language subscribers speak, and contact centers and BPOs can serve globally in more languages while letting customers and agents converse in their native language rather than a shared second one. Three concrete scenarios are described. For live events, translated audio is streamed to attendees' devices at conferences, press briefings and multilingual meetings, so the audience listens in their own language instead of following captions. For telecom operators, live translation can be offered as a feature on the network, so conversations happen seamlessly regardless of what language subscribers speak — the site's example is a Spanish-speaking customer raising a double-charged subscription and requesting a refund. For contact centers and BPOs, customers and agents converse in their native language, extending global service into more languages without requiring agents to speak each one. Streamed translated text and source transcripts can additionally feed live captions or logging in these same deployments. ElevenLabs positions the product as enterprise ready, citing SOC 2, PCI DSS Level 1, ISO 27001, HIPAA and GDPR compliance together with Zero Retention Mode. The audience implied by the site is organisations running high volumes of live conversation across languages: carriers and telecom operators, contact centres and business process outsourcers, and event producers or teams running multilingual meetings and briefings. The site routes commercial enquiries through a contact sales form and offers sign-up and log-in for the ElevenLabs platform; no published price tiers or plan details appear in the provided content, and the capability is documented through the API reference rather than a self-serve pricing table. In short, ElevenLabs Realtime Translation is a streaming, API-driven voice layer that converts live speech into expressive translated audio in 90+ languages, keeps one consistent voice, streams captions and transcripts, and lets teams tune latency against quality per session. Its value proposition is straightforward: a conversation in two languages that sounds like a conversation in one, backed by enterprise-grade security and telecom-ready integration.