DeepL has updated its live translation service with a voice model called Aure that attempts to preserve more than the speaker’s words. The system carries tone, emotion, emphasis and speaking speed into the translated audio, allowing delivery to change as a conversation moves between seriousness, excitement or sarcasm.
Earlier versions created translated speech from a static voice sample captured near the beginning of a conversation. Aure instead analyzes the original audio and translated text continuously. During training, DeepL separated voice identity, speech melody and content so those components could be transferred independently, while reinforcement learning was used to reduce the transfer of accents.
The model now runs automatically in DeepL Voice for online meetings, in-person conversations and the company’s API. It supports English, German, Spanish, French, Italian, Japanese, Portuguese and Korean at launch. Preserving vocal cues could make multilingual meetings easier to follow, but subtle emotion and sarcasm remain context-dependent. Users should not assume that a natural-sounding translation has captured every intention correctly.