DeepL wants real-time translation to finally capture sarcasm and mood shifts

Why it matters

DeepL has introduced Aure, a model designed to make real-time spoken translation sound more natural by carrying over tone, emotion, emphasis, and speaking pace from the original speaker into the target language. Previous versions of DeepL Voice relied on a static voice sample captured when a conversation began, which meant the output could not reflect mood shifts or sarcasm as a discussion evolved. Aure instead continuously analyzes both the translated text and the original audio while the conversation unfolds. During training, DeepL separates voice identity, speech melody, and content so each element can be moved into the new language independently, and reinforcement learning is used to keep accents out of the output. The model runs automatically across DeepL Voice for Online Meetings, In-Person Conversations, and the DeepL API, supporting languages including English, German, Spanish, French, Italian, Japanese, Portuguese, and Korean. The release follows a period of restructuring at DeepL, including job cuts tied to a shift toward smaller AI-focused teams and the acquisition of Mixhalo's team, while Block pursues similar goals with its Dessa speech model.