DeepL Unveils Aure Speech Model to Preserve Tone and Emotion in Real Time

DeepL has launched Aure, a speech translation model that powers its Voice real-time translation product. Unlike conventional systems that only preserve voice timbre, Aure reproduces the speaker's tone, emotional shifts, stress, and speaking rate in the translated output, even capturing sarcasm to make cross-lingual conversations feel more natural.

Aure decouples voiceprint, prosody, and content during training and uses reinforcement learning to suppress accents, ensuring natural-sounding translations while retaining original vocal traits. The model is now automatically available in DeepL Voice's web meeting version, face-to-face mode, and speech API, supporting eight languages: English, German, Spanish, French, Italian, Japanese, Portuguese, and Korean. Chinese is not yet supported, and unsupported languages fall back to the previous model.

The launch marks a shift from mere intelligibility to emotional fidelity in real-time speech translation, enhancing immersion and accuracy for online meetings, cross-border customer service, and international collaboration. By offering Aure via API, DeepL invites developers to build richer voice applications, strengthening its position in the machine translation market.