DeepL Aure carries changing expression into translated speech
German language-AI company DeepL has introduced Aure, a voice-translation model designed to carry changes in tone, emphasis and pace into another language. The model continuously processes source audio alongside translated words, rather than relying on an opening voice sample. It extends the company’s voice-preservation technology across online meetings, face-to-face conversations and its voice programming interface.
Artificial Intelligence··Night
Translation follows a changing voice
German language-AI company DeepL announced Aure on 11 November, extending its real-time speech-translation system to follow changes in a speaker’s delivery. The model powers its real-time voice service and is designed to carry tone, emotion, emphasis and speaking pace into the target language. DeepL says it continuously uses incoming audio together with translated words to determine how the translated speech should sound.[1], [2]
The earlier voice-preservation system worked from a sample taken at the beginning of a conversation. That established a recognizable translated voice, but fixed its expression around the opening sample. Aure changes this processing approach so that later shifts in delivery can shape subsequent translated sentences. DeepL gives changes from seriousness to enthusiasm and urgent emphasis as examples of the expression it aims to retain.[1], [2]
Voice identity, rhythm and content are handled separately
DeepL describes training techniques that separate vocal identity, prosody and spoken content. Prosody is the rhythm and melody of speech, including emphasis and pace. Keeping those components separate allows each to be transferred into the new language independently, according to the company. Reinforcement learning is also used to refine the translated output while retaining speech quality and the meaning of the words.[1], [2]
The system must also distinguish the person’s voice from background noise, echo and audio distortion. DeepL identifies a limited supply of recordings in which the same person speaks different languages as another technical constraint. It says the model adapts to each speaker within a conversation, accounting for changes in delivery as speech arrives rather than requiring a permanent voice sample from every possible situation.[2]
Aure joins meetings and face-to-face translation
The capabilities apply automatically to DeepL Voice for Online Meetings, DeepL Voice for In-Person Conversations and the company’s voice programming interface. Users do not have to configure a separate expression setting. The initial model supports English, German, Spanish, French, Italian, Japanese, Portuguese and Korean. DeepL says its other supported languages continue to use the existing voice models.[1], [2]
Aure builds on the voice-preservation technology that DeepL introduced in September. The company describes its handling of conversation data as part of the technical design: meeting transcripts and translations are not stored permanently, and that information is removed from memory at the end of a session. Adaptation to a speaker therefore takes place during the live exchange, within that stated data-handling arrangement.[2]
