Skip to main content
A Voice is how your PAL sounds. Pick one from the Tavus catalog or create your own from a sample of audio, then reference it by ID from any number of PALs and Faces.
That is the whole configuration. Tavus picks the provider and model that are the best fit for the language(s) of the conversation.

​
Creating a voice

PAL Maker is where you browse the Tavus catalog and create voices of your own, or use the Voices API. To create one, upload a sample of audio and listen to the result in any of the languages Tavus speaks. A new voice can take up to a few minutes to be ready. Training a Face from a video creates a voice from that video’s audio, so the Face speaks in the voice it was trained on. It becomes the Face’s default voice and appears alongside the rest of your voices.
You must hold the rights to the voice in your training audio. See Platform Policies.

​
Using a voice

Set voice_id on a PAL’s TTS layer, and every conversation with that PAL uses it:
Or give a Face a default voice, so any PAL that names the Face needs no TTS configuration at all. Set default_voice_id when you Create Face from an image:
Get Face and List Faces both return two voice references: A Face with no voice set returns null for both rather than omitting them, so you can branch on the value. Change the default with Update Face. That leaves original_voice_id untouched, so setting default_voice_id back to it reverts the Face to the voice it started with. PALs that rely on the default pick up the change on their next conversation.
A voice_id on the PAL takes precedence over the Face’s default_voice_id. Set the default on the Face for the voice that belongs with that likeness, and override on the PAL when one PAL should sound different.

​
Languages

A Tavus Voice sounds natural in every language it speaks, not only the one it was created in. To name the languages to be ready for, set languages on the PAL, or on an individual conversation.

​
Using a voice from your own provider account

Use external_voice_id on the PAL’s TTS layer instead, and manage the voice yourself in your own Cartesia, ElevenLabs, or Azure account. It takes precedence over everything else, including a voice_id on the PAL and a default_voice_id on the Face, and the two voice fields are mutually exclusive on one PAL, so send one. You choose the provider and model, and the same provider voice speaks every language in the conversation, carrying its own accent. A voice_id cannot be combined with your own TTS api_key.