DeepL launches Aure across eight languages for live voice translation

Nathan Brooks

DeepL launched Aure on Nov. 11 with support for eight languages. The model brings changing tone, emotion, emphasis and speaking pace into real-time translated speech.

The useful distinction is delivery, not merely identity. Aure analyzes incoming speech continuously, using both translated words and source audio to decide the delivery of each translated sentence.

In practical terms, Aure resembles a live interpreter more than a prerecorded translation, because it reshapes each translated sentence while the original speaker continues. DeepL, based in Cologne, said the model now powers DeepL Voice, its live voice translation product.

DeepL applies Aure automatically across three offerings, with no setup for users:

• DeepL Voice for Online Meetings
• DeepL Voice for In-Person Conversations
• DeepL API for Voice

How does Aure preserve a changing voice?

Previously, DeepL Voice generated translated speech from a static sample taken during the opening moments of a conversation. The sample captured who spoke, but not how delivery changed, so the model could not follow shifts in emotion or pace.

It also could not convey meaning when delivery contradicted the words, including sarcasm, the company said. Aure was designed specifically to follow those changes as they happen.

Jarek Kutylowski, DeepL's CEO and founder, said: “With Aure, we can capture even the subtlest shifts in expression as they happen and carry them into another language, in someone's own voice and in real time.” The claim covers expression, not merely pronunciation.

DeepL said its training separated voice identity, prosody and spoken content, allowing each to move into the new language independently. The company then used reinforcement learning to maintain translation accuracy and speech quality while removing accents from the output.

Running that process live raises a less tidy problem. Aure has to isolate the speaker from background noise, echo and distortion, and adapt to each voice within a sentence.

That last requirement matters because DeepL deletes meeting audio, transcripts and translations when a session ends. The company therefore has little cross-language data from the same speaker, so Aure must infer delivery during the conversation rather than depend on a long stored history.

What data does DeepL keep?

DeepL says Aure matches the voice live without creating a stored voice profile. It deletes voice data from memory when a session ends, and the company says that data does not train its models.

For buyers, these retention terms sit beside translation performance in any deployment review. They also leave DeepL with little cross-language data from the same speaker, which shapes how Aure adapts within a session.

What changed in November?

Two distinct claims sit behind the launch. Voice preservation, added in September, concerns whether translated speech remains recognizably the speaker's own, while Aure concerns live performance: changing emotion, emphasis, pace and sarcastic delivery.

That distinction matters because preserving a recognizable voice does not by itself preserve the speaker's changing intent. One capability keeps the voice recognizable. The other changes how the translated speech performs.

Which languages and services are available?

Aure initially supports:

• English
• German
• Spanish
• French
• Italian
• Japanese
• Portuguese
• Korean

For every other language, DeepL Voice falls back to its earlier voice models. The broader service therefore runs two generations of voice model at once.

DeepL did not announce separate pricing for Aure, so the launch does not establish whether existing meeting or API prices rise. The sequence began with voice-to-voice translation in April, followed by voice preservation in September and Aure in November.

Who already uses DeepL Voice?

DeepL named Pioneer, Aramark, Avendra International and Inetum as DeepL Voice customers. Its meeting product works in Microsoft Teams, Zoom and Google Meet, while the announcement does not identify which named customers use Aure.

DeepL says that product makes 76% fewer critical translation errors than the platforms' built-in translation tools. The announcement does not define “critical,” identify the sample size or describe the test method, so the figure remains a vendor claim. A useful number, with an asterisk.

For now, buyers outside the eight launch languages will still receive DeepL's earlier models, and DeepL has not said how Aure compares in price with platform-native tools. Those gaps are likely to shape the next phase of enterprise adoption.