The Pulse
DeepL launches Aure voice translation model
DeepL has released Aure across its voice translation products, with initial support for eight languages. The company says the model adapts translated speech to changes in a speaker’s delivery during a conversation.

DeepL has launched Aure, a voice translation model that the company says adapts translated speech to changes in a speaker’s tone, emotion and pace as a conversation unfolds. The model is available in DeepL Voice, its real-time voice translation product.
Source prnewswire.com
Where Aure is available
The model is applied automatically, without additional configuration, in DeepL Voice for Online Meetings, DeepL Voice for In-Person Conversations and the DeepL API for Voice. It initially supports English, German, Spanish, French, Italian, Japanese, Portuguese and Korean, with more languages to follow. Other languages will continue to use existing voice models.
Aure builds on voice-preservation technology designed to produce translations in a recognizable version of the speaker’s voice rather than a generic synthetic one. The company says the new model also follows changes in delivery during a conversation, using the translated words and incoming source audio to shape the spoken output.
From a sample to continuous audio
Previously, the product used the opening moments of a conversation as a static sample to determine the sound of a translated voice. According to DeepL, that approach could capture a speaker’s identity but could not follow later changes in emotion, emphasis or pace, or convey a delivery that contradicted the words, such as sarcasm.
The new model instead analyzes incoming speech continuously. In examples provided by the company, it could reflect a pause before an announcement, emphasis during a negotiation or a shift from a serious tone to enthusiasm when sharing news.
CEO and founder Jarek Kutylowski said the model can capture subtle changes in expression and carry them into another language in the speaker’s own voice in real time. The company also says it can adapt before a speaker has finished a sentence.
Training and data
DeepL says the model must distinguish a speaker’s voice identity and prosody while filtering out background noise, echo and audio distortion. It also identifies transferring a voice across languages without exaggerating accents or distorting how someone sounds as a technical challenge.
To address those challenges, the company combined training techniques that separate voice identity, prosody and spoken content so they can be carried into another language independently. Reinforcement learning further refines the output, preserving translation accuracy and speech quality while producing natural-sounding speech without an accent.
DeepL says it has limited real-world data of the same person speaking across languages. The company also says Voice does not permanently store meeting transcripts or translations and removes that data from memory when a session ends.
Earlier product changes
In September, the company announced voice-preservation models across its voice products and a desktop app for online meetings. Its desktop app for Windows and Mac works with Zoom, Microsoft Teams and Google Meet, and lets users follow conversations through translated subtitles, voice translation or both.
Voice-to-voice translation for online meetings through the desktop app was generally available at that announcement, joining in-person conversations and the voice API. The company said browser-based voice-to-voice translation for external participants remained in beta.