DeepL has updated its voice translation feature with a new model called Aure to preserve speaker tone and emotion during real-time conversations. Previously, the system relied on a static voice sample taken at the start of a session, which prevented it from detecting mood shifts or sarcasm as dialogue progressed. The new approach separates voice identity, speech melody, and content during training to transfer each element independently into the target language. Reinforcement learning is applied to ensure accents do not appear in the output. Aure now operates automatically within DeepL Voice for Online Meetings, In-Person Conversations, and the DeepL API. At launch, the system supports English, German, Spanish, French, Italian, Japanese, Portuguese, and Korean.
This update addresses a core limitation of earlier voice preservation efforts by allowing the system to analyse both translated text and original audio continuously. The change aims to make translated speech sound more natural by carrying over emphasis and speaking pace rather than generating a flat, monotonous voice. This development aligns with broader industry efforts to capture meaning beyond simple word substitution. DeepL recently acquired the team behind Mixhalo to bolster its real-time speech capabilities. Block has also unveiled Dessa, a speech model designed to detect emotions and sarcasm directly from audio.
* DeepL cut approximately 250 jobs in May to restructure around smaller, AI-powered teams.
* The company first rolled out basic voice preservation in September of last year.
* Aure runs automatically without requiring manual configuration from the user.
