Published

DeepL unveils Aure, a voice translation model that carries a speaker’s tone, emotion and pace into real-time translations

On 11 November DeepL unveiled Aure, a Voice AI model powering its real-time voice translation product DeepL Voice. Aure continuously adapts translated speech as the speaker’s delivery changes, carrying tone, emotion, emphasis and pace into the other language. It applies automatically, with no configuration, across DeepL Voice for online meetings, in-person conversations and the voice API.

DeepL’s earlier voice preservation technology let people speak other languages in their own recognisable voice, but it relied on a fixed sample from the opening moments of a conversation: it captured who was speaking, not how, so it could not follow shifts in emotion, emphasis or pace, or convey delivery that contradicts the words, such as sarcasm. Aure instead continuously analyses both the source audio and the translated words to shape how each sentence is delivered.

DeepL says the challenges included isolating a speaker’s voice identity and prosody in noisy conditions, carrying a voice across languages without exaggerating accents, and limited real data of the same speaker in several languages, since DeepL Voice does not permanently store meeting transcripts or translations and deletes them from memory when a session ends. It combined training techniques that separate voice identity, prosody and spoken content, and used reinforcement learning to keep translation accuracy and speech quality.

Aure initially supports English, German, Spanish, French, Italian, Japanese, Portuguese and Korean, with more languages to follow; other languages remain on DeepL’s existing voice models. The press release does not mention pricing changes.

Sources

  1. DeepL unveils Aure, its most lifelike Voice AI translation model yet — DeepL
Open this story in Shift →