Introduction

DeepL has released Aure, a new closed speech-to-speech translation voice model that now powers DeepL Voice. Older voice-matching systems copy a speaker's voice once from the start of a conversation. Aure keeps listening: it uses both the source audio and the translated words to shape how each translated sentence sounds, so a shift into excitement, hesitation, urgency or sarcasm carries over into the other language in the speaker's own voice. You can't download it. It runs inside DeepL Voice for Online Meetings (Teams, Zoom, Google Meet), DeepL Voice for In-Person Conversations, and the DeepL API for Voice. It turns on automatically, with nothing to configure, in eight languages: English, German, Spanish, French, Italian, Japanese, Portuguese and Korean.

Sources: DeepL's press release (November 11, 2026), the Aure product page with side-by-side audio demos, and the DeepL Voice API docs.

What Shipped

  • Model: Aure, DeepL's voice model for real-time voice translation. Closed. No weights, paper, parameter count or latency figure has been published.
  • Modality: speech in, translated speech out (plus live transcripts and text translations), streamed during a live session.
  • What's new: DeepL's September voice-preservation models (announced September 15, 12 languages) built a translated voice from a static sample taken in the opening moments of a call. That captured who was speaking but not how. Aure re-analyzes delivery continuously, so changes in emotion, emphasis and pace show up in the translated speech as they happen.
  • How it's trained (per DeepL): separate training techniques pull apart voice identity, prosody and spoken content so each can be carried into the new language on its own. Reinforcement learning then tunes the output for translation accuracy and speech quality, and for natural-sounding speech without the speaker's accent.
  • Noise handling: DeepL Voice isolates the main speaker and filters out background noise, other voices and echo before Aure shapes the output. DeepL pitches this for open offices and factory floors.
  • Languages: EN, DE, ES, FR, IT, JA, PT, KO at launch. Other DeepL Voice languages automatically fall back to the earlier voice models.
  • Privacy: Aure builds no stored voice profile. Voice and meeting data are processed in memory, deleted when the session ends, and not used for training, according to DeepL.
  • Where to run it today: the DeepL Voice desktop app for Windows and Mac (meetings), DeepL Voice for In-Person Conversations, or the DeepL API for Voice. Speech-to-text and speech-to-speech on the API require a paid DeepL API subscription.

Numbers DeepL Quotes

  • 76% fewer critical translation errors than the meeting platforms' built-in translation tools (DeepL product-page claim for DeepL Voice for meetings, not an Aure-specific benchmark).
  • 4% vs 17% translation error rate for DeepL Voice vs the average of Teams, Meet and Zoom, from Slator testing that DeepL cited in September. That test predates Aure.
  • No MOS, speaker-similarity or latency numbers have been published for Aure itself yet. Use the product-page demos (five emotions: excitement, reassurance, confidence, hesitation and disappointment, translated from English into German, French and Japanese) as the only first-party listening test for now.

Try It on the API

The Voice API is a WebSocket stream. You create a session over REST, then push 16 kHz PCM chunks and read back transcripts, translations and synthesized audio. DeepL says Aure applies automatically to supported target languages. As of this morning, the API changelog has no new Aure-specific parameter, so you request translated speech the same way as before. To keep the speaker's own voice instead of a preset voice, use speaker match (target_media_voice: "match", added September 11):

Install: pip install requests sounddevice websockets

import requests
r = requests.post(
    "https://api.deepl.com/v3/voice/realtime",
    headers={"Authorization": "DeepL-Auth-Key YOUR_AUTH_KEY"},
    json={
        "source_media_content_type": "audio/pcm;encoding=s16le;rate=16000",
        "target_languages": ["de", "ja"],
        "target_media_languages": ["de"],  # one spoken language per session
        "target_media_content_type": "audio/pcm;encoding=s16le;rate=24000",
        "target_media_voice": "match",     # speaker's own voice, not a preset
    },
)
session = r.json()  # then stream base64 audio chunks over the returned WebSocket URL

The full mic-to-headphones example is in DeepL's Real-Time Voice Quickstart. DeepL warns to use headphones, because translated speech played through speakers gets picked up by the mic and translated again.

Why It Matters

Speech-to-speech translation has mostly solved "sounds like me." The next problem is "sounds like I meant it." A static voice clone keeps the timbre but flattens the delivery, so a pause before bad news or a sarcastic "great" comes out as neutral narration. Aure is DeepL's answer to that, and it ships as the default in a real product, not a research demo. Expressive dubbing and translation are getting crowded: ElevenLabs, Google and the full-duplex agent stacks from this week all compete on expressiveness. Aure stands out because it does this live, per sentence, with no stored voice profile.

Caveats:

  • Closed and API-only. There's nothing to run locally and no published architecture.
  • Every quality claim so far comes from DeepL itself. There's no independent Aure benchmark yet.
  • Translated speech for some Voice API languages comes from external partners, and speaker match only works on DeepL-voiced targets. Check GET /v3/languages?resource=voice for each language.

Conclusion

Aure makes DeepL Voice carry your delivery as well as your voice: tone, emphasis and pace follow you into eight languages in real time, with no setup and no stored voiceprint. If you already run the DeepL Voice API, your supported-language sessions should pick it up automatically. If you're comparing expressive speech translation tools, the eight-language demos on the product page are the place to start listening.

—Anabel ♡