Live voice translation from DeepL can now carry a speaker's sarcasm, urgency and changing mood into another language. The speaker, who usually does not know that language, never hears how it came out.
The old voice was a snapshot
In September, DeepL began translating people in their own voice, set from the first moments of a conversation. The system knew who was speaking but not how they felt, so someone who started a call calmly and ended it angry came out calm all the way through. The new model, Aure, keeps listening to the original audio and to the words it has already translated, and works out how each sentence should sound.
Sarcasm is the hardest case, because the delivery contradicts the words. "Great plan" spoken flatly means the opposite of "great plan" spoken warmly, and a voice fixed at the start of the call cannot tell them apart. DeepL trains the model to separate who is speaking, the melody of the speech and the content. That lets the melody be rebuilt in the new language on its own while the voice stays recognizable as the speaker's.
A refusal is often all tone
Picture a buyer in Munich on a call with a supplier in Osaka, asking whether a delivery can move up a week. In many business cultures a refusal comes as a pause and a slower, softer voice rather than the word no. An even voice reading out the translated words makes that answer sound like a maybe. With the pause and the pace carried across, the buyer hears the no and can offer another date on the same call, instead of finding out from a missed shipment.
Support teams gain the same thing. DeepL also sells its voice translation through an interface for contact centers, and an agent who can hear that a caller is close to hanging up has more to work with than the same words in a calm voice.
The speaker cannot check the result
The buyer in Munich speaks German, so they have no way to judge whether the Japanese voice sounded irritated, polite or sarcastic when they meant none of those. The model builds the delivery from the audio and the words, which will be right most of the time. When it is wrong, a mood the speaker never had goes out in their own voice, and nobody on their side notices.
That is a good trade for most calls, and it still changes some habits. Sarcasm is the first thing to drop when a machine is in the middle, because it is the delivery where a wrong guess turns the meaning around. Anything that commits money or time deserves two plain lines in writing afterward, saying what was agreed.
Expect every meeting tool to compete on this. Microsoft Teams can already copy a speaker's voice with the speaker's consent, and we wrote in October about a voice model that hears the tone text drops. Translation is now working on the same gap. For now the new voice covers English, German, Spanish, French, Italian, Japanese, Portuguese and Korean. A call that includes any other language still gets the older snapshot voice, so mood shifts on that call are still lost.