DeepL Aure explained: voice translation that follows tone

By LLMTest Lab · Nov 11, 2026 · 3 min read voicetranslationdeepl
On this page

On this page

  1. What it is
  2. What is new
  3. The claims, and who made them
  4. Who should care
  5. What we will test

Not tested by us yet. Aure is only reachable through DeepL's own products and its voice API, and our test harness calls chat models, so we have not listened to it ourselves. Everything below comes from DeepL's announcement of Aure. We'll update this post when we can run it on a fixed set of recorded sentences.

Aure is DeepL's new voice model for real-time speech translation, announced on 11 November 2026. It changes the translated voice as the speaker's tone, emotion, emphasis and pace change, instead of reading out a flat translation. It switches on automatically for eight languages inside DeepL Voice and the DeepL API for Voice. If you build meeting or contact-centre tools on DeepL, it arrives with no code change. If you need a price or a latency figure, DeepL has published neither.

DeepL Aure
Maker DeepL
Announced 2026-11-11
Type Real-time speech translation voice model
Launch languages English, German, Spanish, French, Italian, Japanese, Portuguese, Korean
Other languages Keep DeepL Voice's existing voice models
Access DeepL Voice for Online Meetings, DeepL Voice for In-Person Conversations, DeepL API for Voice
Setup None, applied automatically
Price Not stated in the announcement
Latency, model size Not stated

What it is

DeepL describes Aure as its most dynamic and lifelike voice model so far. It builds on the voice preservation feature that DeepL added to DeepL Voice earlier, which kept a speaker's recognisable voice in the translation. Aure adds the delivery on top: an angry sentence or a slow, emphasised one should sound that way in the other language.

According to DeepL, the model analyses speech continuously using both the translated words and the source audio. It filters background noise, echo and distortion, and adapts to each speaker over a session. Training separates voice identity, prosody and content, and reinforcement learning then tunes the output for accuracy and natural, accent-free speech.

What is new

DeepL launched DeepL Voice-to-Voice in April 2026 as a live spoken translation suite. The new part is expressiveness. Earlier voice translation, from DeepL and others, tends to get the words right and flatten the delivery. Aure is DeepL's attempt to carry tone and pacing across.

DeepL also says it does not permanently store meeting transcripts or translations, and that it deletes session data from memory when a session ends. For regulated buyers that is a bigger deal than the voice quality.

The claims, and who made them

Every claim above is DeepL's. The announcement gives no word error rate, no listener scores, no latency and no comparison with other speech-to-speech systems. "Most lifelike" is a marketing description, not a measurement. Treat it as unverified until someone publishes numbers on shared test audio.

Who should care

  • Teams already on DeepL Voice. You get Aure in the eight launch languages without changing anything, so plan a listening test on your own calls before and after.
  • Buyers comparing voice translation vendors. Ask for latency under load, behaviour on noisy audio and per-minute pricing, none of which are public yet.
  • Builders who want to self-host voice. Aure is not an open model. Our look at ZONOS3-Mini, an open text-to-speech model and a 32MB open TTS model on CPU cover the self-hosted end of voice.

Skip it if your languages fall outside the eight, since the rest stay on the older voice models.

What we will test

We would play 20 recorded sentences with clear emotional shifts (a complaint turning polite, a question, a joke) through Aure and the previous DeepL voice, in English to German and English to Japanese. Native listeners would score tone match and naturalness, a speech recognizer would score word accuracy, and we would time the delay from end of phrase to first translated audio. Until then, every quality claim here is a vendor claim.

Related articles

ZONOS3-Mini explained: 100M-active open TTS, 43 languages
Zyphra's ZONOS3-Mini is an Apache 2.0 text-to-speech model with 800M total and 100M active parameters, voice cloning and a browser build. Specs and claims.
Underdog Woof first look: a 32MB open TTS model on CPU
Underdog Woof is a 38M-parameter English text-to-speech model that runs on CPU in a 31.7 MB download. Specs, vendor-reported WER, license limits, who it suits.