DeepL adds Aure voice model that carries emotion into translation

DeepL's previous voice models worked from a snapshot: they captured the speaker's voice at the beginning of a conversation and generated translated audio from that baseline. Aure shifts to continuous analysis, using both the translated words and the original audio to track how delivery evolves mid-conversation. The company points to situations like a negotiation where emphasis conveys urgency, or good news delivered with a shift from a serious tone to an enthusiastic one, as cases where the added nuance matters.

To build this, DeepL combined multiple learning methods that separate a speaker's distinctive vocal traits, intonation and rhythm from the content of what is said. It then used reinforcement learning to optimize the output, aiming to keep translation accuracy and audio quality while producing natural speech that does not favor a particular accent.

The rollout is designed to be invisible to users: Aure applies automatically across DeepL Voice for Online Meetings, DeepL Voice for In-Person Conversations and DeepL API for Voice, with no extra setup. Support currently covers 8 languages — English, German, Spanish, French, Italian, Japanese, Portuguese and Korean — with expansion planned.