DeepL launches Aure voice model that carries tone, emotion and speaking pace into translated speech

DeepL's new Aure model carries tone, emotion, emphasis and speaking pace from the original speaker's voice into the translated language, the company announced on Nov. 11, 2026. It runs automatically in DeepL Voice for Online Meetings, In-Person Conversations and the DeepL API, according to The Decoder's report.

What changed

Until now, DeepL Voice generated translated speech from a static voice sample captured at the start of a conversation. That design meant the system could not pick up on mood shifts or sarcasm during a conversation.

Aure works differently. It continuously analyzes both the translated text and the original audio as the conversation unfolds, so the output can follow changes in delivery rather than reuse a fixed sample.

How it is trained

During training, DeepL separates voice identity, speech melody and content. Each component can then be transferred into the new language independently. Reinforcement learning is used to keep accents out of the output.

Availability and language support

Aure is active by default in three places:

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

At launch, the model supports eight languages: English, German, Spanish, French, Italian, Japanese, Portuguese and Korean.

Pricing, parameter count, context window and benchmark results have not yet been disclosed.

Background

DeepL first rolled out voice preservation in September. In May, the company cut about 250 jobs to restructure around smaller, AI-powered teams, part of what it called becoming an "AI-native" company. DeepL has also acquired the team behind audio streaming specialist Mixhalo to strengthen its real-time speech translation.

Block is pursuing a similar goal of capturing meaning beyond words. The company recently unveiled Dessa, a speech model designed to detect emotions and sarcasm directly from audio.

What this means

The key change is architectural. A voice sample captured once at the start of a conversation cannot reflect how a speaker's mood changes. By analyzing the original audio continuously, Aure can reflect those changes in the translated speech, including sarcasm, which the earlier approach could not handle.

Separating voice identity, melody and content during training is what allows each element to be carried into a different language independently. The reinforcement learning step targets a specific failure mode, accent leakage, so the translated voice does not inherit the source language's accent.

Because Aure is on by default across Online Meetings, In-Person Conversations and the API, existing DeepL Voice users and API developers get the change without a separate integration. The eight-language launch set defines the current scope.