DeepL has introduced its new Aure model, designed to enhance real-time translation by preserving the speaker's original intonation, emotion, and pace. Unlike previous versions that relied on static voice samples, Aure continuously analyzes both the translated text and the original audio stream as a conversation progresses.
The training process for Aure involves separating voice identity, speech melody, and content. This allows the system to transfer these elements independently into the target language while using reinforcement learning to minimize foreign accents. The model is now active across DeepL Voice for Online Meetings, In-Person Conversations, and the DeepL API.
At launch, the technology supports eight languages: English, German, Spanish, French, Italian, Japanese, Portuguese, and Korean. This release follows the initial rollout of voice preservation features in September. The company continues to focus on AI-native development, including the acquisition of the team behind audio streaming specialist Mixhalo.
