DeepL has a new model called Aure that makes its live translated speech sound more like the person talking. It carries over tone, emotion, emphasis and speaking pace into the other language. The goal is for sarcasm and mood shifts to come through, not just the words.
What changed
Until now, DeepL Voice built translated speech from one voice sample taken at the start of a conversation. That meant it could not follow mood changes or sarcasm. Aure keeps analyzing both the translated text and the original audio as the conversation goes on. DeepL also uses reinforcement learning, a training method that rewards good results, to keep accents out of the output.
Aure runs automatically in DeepL Voice for Online Meetings, In-Person Conversations and the DeepL API, which lets other software connect to DeepL. At launch it supports English, German, Spanish, French, Italian, Japanese, Portuguese and Korean.
What is not known yet
The material does not give a price, a rollout date or any test results. It is unclear how well Aure handles sarcasm in real calls. DeepL first rolled out voice preservation in September, so this builds on that. Block is chasing a similar goal with Dessa, a model that detects emotions and sarcasm from audio.
The takeawayIf you run calls in other languages, watch for Aure in DeepL Voice, but test how it handles tone before relying on it.