DeepL's Aure Keeps Tracking Tone and Pace Mid-Sentence So Translated Speech Still Sounds Like You

DeepL has released Aure, a new Voice AI model that tries to carry not only a speaker’s identity but also how that voice changes as a conversation unfolds. The Cologne company announced the model on November 11, 2026. Aure powers DeepL Voice, DeepL’s real-time voice translation product, and the company says the new behavior is applied automatically across DeepL Voice for Online Meetings, DeepL Voice for In-Person Conversations and the DeepL API for Voice.

The launch follows DeepL’s earlier voice-preservation work, which aimed to keep a recognizable personal voice in the translated audio rather than a generic synthetic one. Aure’s claim is the next step: follow shifts in tone, emotion, emphasis and pace as they happen, and reflect them in the translated speech in the speaker’s own voice.

From a fixed snapshot to continuous listening

DeepL says previous DeepL Voice models could generate a natural, expressive voice from a static sample taken from the opening moments of a conversation. That captured who was speaking, but not how they were speaking later. A fixed snapshot cannot follow a change in emotion mid-call, a sudden emphasis, a slower pace or delivery that contradicts the words, such as sarcasm.

Aure, DeepL says, analyzes incoming speech continuously, using both the translated words and the source audio to shape how each sentence is delivered. The company frames several hard problems that sit under that sentence. Real conversations are noisy, so the model has to isolate the speaker’s identity and prosody while filtering background noise, echo and distortion. Carrying a voice across languages is harder than within one language, especially without exaggerating accents or distorting how someone sounds. And DeepL Voice does not permanently store meeting transcripts or translations; it deletes that data from memory when a session ends. Aure therefore has to adapt to each speaker in real time, before they have finished a sentence, including day-to-day variation such as sounding tired, enthusiastic or stressed.

To do that, DeepL says it combined training techniques that separate a speaker’s voice identity, prosody and spoken content so each can be carried into a new language independently. Reinforcement learning then refines the output for translation accuracy and speech quality while aiming for natural-sounding, accent-free speech. Those are the company’s own descriptions of its method; DeepL did not publish a model size, latency distribution or third-party benchmark scores with the press release.

Eight languages at launch, more later

At launch, Aure supports English, German, Spanish, French, Italian, Japanese, Portuguese and Korean, with more languages to follow. DeepL Voice will keep using its existing voice models for other languages. DeepL says no additional configuration is required for Aure’s capabilities.

The DeepL Voice API, documented separately for developers, streams audio over a WebSocket and can return source transcripts, translations into multiple target languages and translated speech within one session. Speech-to-text and speech-to-speech are available to customers with a paid DeepL API subscription. The Voice language matrix lists many more languages for transcription and translation than Aure currently covers for its new prosody behavior; developers who need the full matrix still use the existing Voice stack for languages outside Aure’s initial eight.

Why DeepL thinks expression is the product

CEO and founder Jarek Kutylowski cast the release as reversing a long compromise in cross-language talk: people have had to change how they speak to be understood, rather than the system adapting to them. “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,” he said in the announcement.

DeepL points to enterprise users of DeepL Voice including Pioneer Corporation, Aramark and Avendra International, and Inetum, and notes that DeepL Voice was recognized in Fast Company’s 2026 Next Big Things in Tech list in the Applied AI category.

Aure is not an open text-to-speech checkpoint but a closed model inside a paid product, and its central claim, that tone and pace survive the change of language, can only be judged in use. Until DeepL or independent testers publish latency and quality measurements, buyers have the eight-language list and the automatic rollout to go on, and little else.