DeepL Voice adds live speech-to-speech translation

DeepL Voice was first announced in November 2024 as a real-time voice translation service, split into DeepL Voice for Meetings for web conferences and DeepL Voice for Conversations for in-person talks. Until now it translated speech into text; the speech-to-speech feature was shown as in development in October 2025 and ran as a beta with early access from April 2026. General availability arrives alongside a new Voice AI model that carries tempo, intonation, pitch and tone into the translated audio.

On the meetings side, DeepL is addressing an old friction point. In the browser version, users had to paste a meeting link into DeepL and manually send a translation bot into the call. The new Windows and Mac desktop app detects Zoom, Microsoft Teams and Google Meet sessions and connects them with one click. For in-person use, the group conversation feature creates a virtual room where participants join from their own devices, up to 20 at a time, rather than passing a single smartphone around.

DeepL is also positioning itself inside other AI tools. Its MCP support, announced on the 22nd, lets Microsoft Copilot, ChatGPT and Claude call DeepL for translation with a company's own glossary and custom rules, without users copying text out and pasting results back. On hardware, DeepL says it has no plans to build AI glasses itself, but its voice translation API leaves room for device makers to build translation display apps on top of it.

The immediate open question is latency. DeepL Voice currently runs speech recognition, translation and voice output as three separate models in sequence, and DeepL is developing a unified model that handles input to output in one pass. Whether that unified model lands will likely shape how close the translated speech feels to natural back-and-forth conversation.