Translated speech can convey more than the words someone says. DeepL’s new Aure model is designed to carry features such as tone, emotion, emphasis, and speaking pace into another language as a conversation unfolds.
From a fixed voice sample to changing speech
Previously, DeepL Voice generated translated speech using a static voice sample captured at the start of a conversation. That approach could preserve a speaker’s voice, but it could not respond to changes in mood or pick up on sarcasm as the conversation continued.
Aure takes a different approach. It continuously analyzes both the translated text and the original audio, rather than relying only on the initial sample. That allows it to account for how a speaker’s delivery changes while preserving aspects of the original performance in the translated speech.
The distinction matters because the same words can carry different meaning depending on how they are said. Emphasis, speech melody, and pace can help listeners interpret what a speaker intends. DeepL’s stated goal is to retain those cues across languages, including when a speaker’s tone shifts or becomes sarcastic.
How Aure handles voice and language
During training, DeepL separates voice identity, speech melody, and content. The company says each component can then be transferred into the new language independently. This setup gives the model separate elements to work with: whose voice is being represented, how the speech sounds, and what is being said.
Reinforcement learning is used to keep accents out of the output. Together, these choices are intended to let translated speech carry the original speaker’s vocal expression without simply reproducing every feature of their accent.
Aure runs automatically in DeepL Voice for Online Meetings, In-Person Conversations, and the DeepL API. The source says the model supports English, German, Spanish, French, Italian, Japanese, Portuguese, and Korean at launch.
Why preserving delivery matters
Real-time translation has to handle spoken communication as it happens. A translation that conveys the text but loses the speaker’s emphasis or pace may leave out information listeners use to understand the exchange. By analyzing audio and text continuously, Aure is built to keep more of that delivery in the translated voice.
That does not mean every subtlety is guaranteed to transfer perfectly. The source describes the model’s design and supported languages, but does not provide performance measurements or examples showing how reliably it recognizes sarcasm or emotion. Its stated capability is to carry vocal qualities into translated speech as conversations progress.
The approach also reflects a broader interest in extracting meaning from audio itself. The article notes that Block recently unveiled Dessa, a speech model designed to detect emotions and sarcasm directly from audio. Aure’s described role is different: it carries vocal qualities into translation rather than being presented as a tool for detecting those qualities on their own.
Part of a wider push into speech translation
DeepL first rolled out voice preservation in September. The article also reports that, in May, the company cut about 250 jobs as it restructured around smaller, AI-powered teams, a change it described as becoming an “AI-native” company. DeepL acquired the team behind audio streaming specialist Mixhalo to bolster its real-time speech translation.
Aure brings those strands together in a feature that runs within DeepL Voice and its API. Its focus is not only on converting spoken content into another language, but also on retaining aspects of how the original speaker delivered it. For people following a translated conversation, that could make changes in emphasis, pace, or mood easier to hear.