Skip to main content
Set languages on a PAL to name the languages its conversations are expected to run in, and override it on an individual conversation when that one needs a different set.
This makes recognition more accurate, keeps the PAL from answering in a language you did not ask for, and picks the right voice for each language automatically.
properties.language is deprecated. It still works and existing integrations are unaffected, but it only names one language and cannot say what a conversation should be prepared for. Prefer languages, including for a single language: languages: ["es"] replaces language: "spanish". See Migrating from language.

Setting languages

cURL
What it accepts:
  • Language codes from the spoken languages below.
  • An English accent tag in place of en.
  • Entries in order. The first entry is the language the conversation opens in.
Name only the languages you expect. A narrower set is more accurate in those languages than a wide one, so declaring five beats declaring twenty when you only need five. With languages set on neither the PAL nor the conversation, the conversation runs in English, the same as sending ["en"]. If no speech-recognition engine supports a language in languages, Tavus rejects the conversation with a 400. A conversation whose pinned STT or TTS engine cannot support every language in the field is also rejected. If you pin an engine or a model, check that it covers the languages you declare.

PAL and conversation precedence

Every conversation with a PAL starts from that PAL’s languages. Setting languages on a conversation applies to that conversation alone.
Omit languages on the conversation to inherit the PAL’s. Sending it replaces the PAL’s set. The two are not merged.
Both fields accept exactly the same values.

English accents

To give the voice an English accent, put one of these tags in languages in place of en, at the same position. The first entry stays the language the conversation opens in.
Set the tag on the PAL’s languages, or on properties.languages for one conversation, with the same precedence. There is no separate accent field. In PAL Maker, the same choice is under Advanced settings → Languages → English accent.
  • Voice: a Tavus Voice plays the accent, as the PAL’s voice_id or as the face’s default_voice_id when the PAL sets none. A voice from your own provider account (external_voice_id) keeps its own accent.
  • TTS engine: leave tts_engine unset, or set it to tavus-auto or cartesia. With any other engine, the tag does not change the accent.
  • One tag per language: if languages holds two English tags, Tavus uses the first one.
  • Casing: writes accept any case. A PAL reads back with tags in display case, so en-gb comes back as en-GB.
Only English has accent tags. There is no pt-br, es-mx, or fr-ca accent. For a native accent in another language, create a voice from your own audio recorded in that accent, or use a voice from your own provider account.

Spoken languages

Tavus supports 42 languages with the default tavus-auto TTS engine, alone or in any combination.
Georgian (ka) works only on its own. Set languages: ["ka"] for a Georgian conversation. In a list with other languages, Tavus skips Georgian, and the other languages work as usual.
tavus-auto picks the provider and model for you. Set tts_engine and tts_model_name explicitly only when you need a specific one, and check its language coverage first.

Migrating from language

Additional language support via Azure

Azure can speak some languages outside the 42 above. It replaces only the voice: speech recognition must still support the language, or the conversation is rejected with a 400 (see Setting languages). Use Azure only as a fallback, prefer the default tavus-auto routing whenever your language is already supported. For the latest list, see Azure Speech TTS language support. Unlike Cartesia and ElevenLabs (which Tavus hosts for you), Azure requires you to bring your own Azure Speech API key:
  1. Set tts_engine to azure explicitly on the PAL’s TTS layer.
  2. Provide your own Azure Speech resource api_key. The resource must be in the East US region (Tavus synthesizes via eastus; a key from another region returns an authentication error). Set external_voice_id to any standard Azure neural voice available in East US (for example de-DE-KatjaNeural); Custom Neural Voices must be deployed in that same resource.
Any Azure voice can speak any supported language, carrying the voice’s own accent. For a natural accent, pick a voice whose locale matches your target language (e.g. a de-DE-* voice for German), or use one of Azure’s *MultilingualNeural voices.

Per-engine language lists

Cartesia

ElevenLabs