Skip to content

AI models and routes

The right route per language. Not one model for everything.

Pick a route per language. Tõlk keeps it running and switches engines if one fails. Availability differs by product and language.

Direct speech-to-speech

A realtime model listens and speaks the translation directly, keeping the speaker’s pace.

  1. Speech in
  2. Realtime translation model
  3. Translated speech out

Best for

  • Keynotes and talks
  • Fast, natural delivery
  • Many languages at once

Trade-offs

  • Names and jargon follow the model’s own understanding
  • Voice is the model’s own

Engines and voices

  • OpenAI realtime translation · Default engineAvailable
  • Gemini Live Translate · Widest language listAvailable
  • Qwen LiveTranslate · Opt-in; processed outside the EEAPilot/beta
Conference Spotlight Polylogue Bridge

Dictionary-assisted

Speech is recognized, translated with your event Dictionary, then read by a chosen voice.

  1. Speech in
  2. Recognition
  3. Translation with your terms
  4. Voice

Best for

  • Names, brands and terminology
  • Consistent voices per speaker
  • Specialist seminars

Trade-offs

  • Slightly more delay than direct translation
  • Works sentence by sentence

Engines and voices

  • OpenAI recognition and translationAvailable
  • OpenAI or ElevenLabs voicesAvailable
Conference Spotlight Bridge

Prepared and reviewed

For recordings: translate and voice ahead of time, review, then broadcast or export.

  1. Recording
  2. Transcribe + translate
  3. Voice per speaker
  4. Human review

Best for

  • Recorded talks and courses
  • Scheduled premieres
  • Content you will reuse

Trade-offs

  • Not live: needs preparation time

Engines and voices

  • Standard voicesPilot/beta
  • Premium studio voicesPilot/beta
Studio

How a route is chosen

What matters for your content.

Language coverage

Engines support different language lists. Tõlk shows which engines can serve each language.

Delay

Direct routes start speaking sooner; dictionary routes trade a little time for precision.

Terminology

When names and terms must be exact, a Dictionary-assisted route uses your glossary.

Voice

The model’s own voice, or a chosen voice per speaker.

Region and privacy

Some engines process outside the EEA; they are opt-in and labelled.

Resilience

If an engine fails, languages switch to a backup automatically and switch back on recovery.

Voices

Studio-voice and dictionary routes can read translations in OpenAI voices (Marin, Cedar, Coral, Sage, Ash, Alloy, Ballad, Echo, Fable, Nova, Onyx, Shimmer, Verse) or ElevenLabs voices (George, Sarah, Daniel, Matilda, Brian, Alice), set per language, and in some products per speaker.

Technical documentationTest routes on your content

The Live AI Translation Hub

Ready for your next moment?

Tell us what to make multilingual. We pilot it on your content, then you decide.

Request a pilotTalk to the team · hello@tolk.fm