DeepL Voice models

How the voice family differs from text translation

DeepL's researchers first adapted its text translation model for faster inference. They then developed voice models that decide when enough speech is available to produce a translation. This work aims to reduce delay and repeated subtitle corrections without waiting for every sentence to finish.DeepL AI Labs: from text models to real-time voice

The same research page describes direct speech-to-speech models without an intermediate text stage as work in development. It does not establish that every released Voice route uses one end-to-end audio model.DeepL AI Labs: from text models to real-time voice

DeepL and partner language routes

DeepL provides recognition and speech output for languages including English, German, French, Spanish, Japanese and Korean. The table marks other routes as externally supplied. For example, Arabic recognition and translated speech use a service partner. Some languages have translated text but no translated speech output.DeepL Voice languages and external providers

External recognition routes require an explicit source language because automatic detection is not available for them. The API's languages response exposes an external flag for recognition and speech output. Check each stage when the choice of service provider matters to an integration.DeepL Voice languages and external providers

Live voice preservation

The September 2026 release introduced models that retain a speaker's voice, rhythm and expression across languages. DeepL states that this happens live without storing voice samples or building a reusable voice profile. The release covers online meetings, in-person conversations and the Voice API.DeepL September 2026 voice preservation release

The desktop application supports Zoom, Teams and Google Meet, with captions and spoken translation. Its voice-to-voice feature is generally available. The release still labels browser-based voice-to-voice access for external participants as beta. That product limit does not make the whole model family a preview.DeepL September 2026 voice preservation release

Streaming limits and customization

The API accepts a single-channel stream and limits input speed to twice real time. Audio chunks should stay below 100 kilobytes or one second. DeepL recommends chunks lasting 50 to 250 milliseconds. No input for 30 seconds ends the session, and clients must handle interrupted connections.DeepL Voice API audio and session requirements

Spoken Terms influence how names and specialist vocabulary are transcribed. They are monolingual term lists, with at most 300 characters per language list. Glossaries control the translated form of a term. These settings address different stages and should be configured separately.DeepL Spoken Terms guide

Streaming glossaries match the transcription as it arrives. A long multi-word term can be missed if its words are completed in separate segments. Glossary coverage therefore needs review in live speech, even when the same dictionary works for a complete text.DeepL Voice API overview

Practical assessment

Practical guidance. Review Voice with the microphones, accents and turn-taking patterns that the application will use. Check the original transcript before judging the translation: a wrong product name can begin at recognition. Then check how long listeners wait for usable meaning and whether subtitles change after they have been read.DeepL AI Labs: from text models to real-time voiceDeepL Spoken Terms guide

Practical guidance. If two listeners need different spoken target languages, the documented one-target limit means the application must manage separate sessions. Account for this when designing the conversation flow. Expose the selected input and output languages so each participant knows what the service will translate.DeepL Voice API audio and session requirements