DeepL Aure translates sarcasm, cheers and whispers in real time

What's new

On November 11, 2026, Cologne-based AI company DeepL unveiled Aure, a new speech AI model for live translation of spoken language. DeepL calls it the most lifelike speech model it has built so far. Aure doesn't just translate what you say but also how you say it. Tone, emotion, emphasis and speaking pace are meant to carry over into the translated speech, and you still hear the speaker's own voice.

South Korean trade paper Digital Daily explains it with a football example. "Ah, goal!" sounds completely different depending on who scored. If it's your team, it's a cheer. If it's the other side, it's a groan. Aure is supposed to carry exactly that kind of difference into the target language.

Aure powers DeepL's real-time translation product DeepL Voice, and you don't have to set anything up to use it. It runs in DeepL Voice for Online Meetings, in DeepL Voice for In-Person Conversations and in the DeepL API for Voice, which companies use to build the technology into their own applications. At launch, Aure supports eight languages: German, English, Spanish, French, Italian, Japanese, Portuguese and Korean. More are planned. For all other languages, DeepL Voice keeps using its existing speech models.

What was missing

Aure builds on the voice preservation feature DeepL launched recently. With it, the people you talk to hear the translation in your recognizable voice rather than a generic computer voice. DeepL says that approach had a weakness: the translated voice was built from a fixed voice sample captured in the first moments of a conversation. That told the system who was speaking, but not how. In DeepL's words:

A fixed snapshot can't follow shifts in emotion, emphasis or pace, and it can't convey a delivery that contradicts the words – sarcasm, for example.

Aure, by contrast, analyzes incoming speech continuously, using the translated words and the original audio at the same time. If someone gets louder or speaks faster, the translation changes to match. Pause briefly before a big announcement, push hard in a negotiation, or go from serious to excited while sharing good news, and that should come across in the target language too.

How Aure works under the hood

The key difference is the architecture. When DeepL launched Voice-to-Voice in April 2026, TechCrunch and The Next Web reported that the system still worked as a three-step pipeline. Speech was converted to text, the text was translated with DeepL's translation model, and the result was turned back into speech. Back then DeepL already said it planned to build an end-to-end model that skips the text step. Aure is that model. Speaking to Korean journalists, including Digital Daily, in San Francisco in late October, DeepL's head of engineering Vik Singh said:

Earlier models first converted speech into text and then translated it. The new model takes in speech directly and generates translated speech. That lets us bring whispers, sarcasm and the information in intonation into the translation.

DeepL points to several technical challenges. Real conversations are noisy, so Aure has to separate the speaker's voice identity and prosody (intonation, rhythm and emphasis) from background noise, reverb and distortion. Carrying a voice into another language is also much harder than working within a single language, especially if you don't want to exaggerate accents or distort how the person sounds.

Then there's a data problem, and DeepL accepts it on purpose. DeepL Voice doesn't store transcripts or translations permanently and wipes them from memory when a session ends. As a result, there are hardly any real recordings of the same person speaking several languages for the model to learn from. Aure has to adapt to each speaker in real time, before they've even finished their first sentence. It's also supposed to pick up how someone is doing that day, whether they sound tired, excited or stressed.

To get there, DeepL combines several training techniques that separate voice identity, prosody and content, so each can be carried into the new language on its own. Reinforcement learning (training a model by rewarding good results) then refines the output further. The goal is accurate translation, high audio quality and speech that sounds natural and accent-free. According to Singh, DeepL doesn't just train on evenly spoken news or podcasts. It also uses emotionally charged sports broadcasts and call center conversations with patchy audio quality.

Background: DeepL Voice since April

DeepL is best known for text translation and was founded in 2017 by CEO Jarek Kutylowski. The company says it has more than 900 employees and over 200,000 business customers. In April 2026 it launched the DeepL Voice-to-Voice product line. It included translation in Microsoft Teams and Zoom, conversations on mobile and web without installing an app, group conversations you join via QR code, and the API, aimed at uses like call centers. A feature called "Spoken Terms" also lets you add technical terms, company names and people's names.

In a blind study by industry analyst Slator, commissioned by DeepL, 96 percent of professional linguists preferred DeepL Voice over the built-in translation in Google Meet, Microsoft Teams and Zoom. According to The Next Web, though, a live demo in Seoul in April showed a lag of one to two sentences. Chief product officer Gonzalo Gaiolas put that down to differences in word order and sentence structure between languages. DeepL's latest announcement doesn't say whether Aure reduces this latency, the delay between what's said and the translation.

DeepL lists Pioneer Corporation, Aramark, Avendra International and Inetum among its DeepL Voice customers. Fast Company put DeepL Voice on its "Next Big Things in Tech 2026" list in the Applied AI category.

Our take

Aure is a big step for DeepL. It moves the company away from a chain of speech recognition, text translation and speech synthesis toward a model that processes speech directly. In business conversations especially, that can matter a lot: whether a sentence is meant seriously, ironically or enthusiastically changes what it means. If you use DeepL Voice, you get the upgrade automatically, with nothing to switch on.

The market is crowded, though. Google, Microsoft and Zoom have their own translation features built into their platforms, and startups like Palabra (backed by Alexis Ohanian's Seven Seven Six) are also working on real-time translation that keeps your voice. Sanas and Camb.AI work in adjacent areas such as accent conversion and dubbing.

Some questions remain open. DeepL hasn't published any benchmarks for Aure, either for translation quality or for latency, and there are no independent tests yet. It remains to be seen whether end-to-end processing can match the translation quality of DeepL's proven text models, which the company was still explicitly relying on in April. At launch, Aure covers only eight languages, while DeepL Voice supports more than 40 overall. And the more convincingly an AI can reproduce your voice, emotions included, the more pressing the question of misuse through voice fakes becomes. DeepL's announcement doesn't address that.