Voices
Fetch the live list at runtime instead of hardcoding:
Languages
Versa speaks nine languages.
Every voice can speak every one of these: the model is zero-shot, so a voice is a timbre rather
than a language, whatever language its reference speaker was recorded in.
Fetch the live list at runtime rather than hardcoding it:
Which languages are understood
These nine are the languages a voice can speak. Speech recognition, which voice agents and the speech-to-text API use to understand a caller, covers six of them: English, Spanish, Portuguese, French, Italian and German. Catalan, Galician and Basque are speak-only for now: text-to-speech and an agent’s voice use them, but a caller speaking one of them is not understood.Models
versa-1.0 has been retired and is no longer served.
Parameters
Speed
Accepted for compatibility on every endpoint, but has no effect onversa-2.0: the model gets its
timing from its own duration predictor, and there is no rate control in the current graph. Sending
a value neither errors nor changes anything.
Steps / quality
steps (internally nfe, the number of denoising steps) is accepted everywhere but has no effect on the
current voice model. Since 2026-09-29, versa-2.0 serves a distilled model that runs a fixed 4 steps on
every request: voice agents, WebSocket streams and REST alike. It was trained for exactly that schedule, and at
it the model is faster and makes fewer word errors than the previous one did at 6.
Validation is unchanged, so a value outside the accepted range still returns
validation.steps_out_of_range.
The deprecated quality alias on REST is accepted the same way.
Voice character
cfg (1–4 on REST, per agent and per session) is accepted but has no effect on the current voice model,
which is trained to run without guidance: the voice comes from the reference alone. Until 2026-09-29 the default
was 2, with projected guidance.