DeepL's New Aure Model Brings Tone and Emotion to Live Voice Translation

At a glance

DeepL has added a new voice model called Aure to its real-time interpreting service, DeepL Voice. Before, the service made translated speech from a short voice sample taken at the start of a conversation, so it missed mood changes and sarcasm. Aure keeps analyzing the translated text and the original audio as the conversation goes on.

Key points

1. 1

   Aure carries the speaker's tone, emotion, emphasis, and speaking speed into the translated speech.

2. 2

   During training, it separates voice identity, speech melody, and content so each can move into the new language.

3. 3

   Reinforcement learning, a method that rewards good outputs, keeps accents out of the translated speech.

4. 4

   It runs automatically in online meetings, in-person conversations, and the DeepL API, and supports English, German, Spanish, French, Italian, Japanese, Portuguese, and Korean.

Why it matters

Translation that only carries words over can easily flip the meaning of a joke or a sarcastic remark. This technology narrows that gap, showing that voice translation is moving closer to real conversation.