Voices
A voice is chosen by slot: five female (F1–F5) and five male (M1–M5). Each language has its own native speaker in each slot, so the speaker you hear depends on language as well as voice: F1 in Spanish is a Spanish speaker and F1 in English is an English speaker, not the same person speaking two languages. Keep voice and language together when you want a consistent result.
GET /voices lists the slots with these display names:
The names label the slot; they are the same whatever
language you use. Where a language has no speaker of its own in a slot, the nearest slot of the same gender is used: Spanish F2 sounds as F1, Italian F3 as F2, Portuguese M5 as M4, and French M3–M5 as M2. English, Spanish, Catalan, Galician, Basque, Portuguese, French, Italian and German each have their own speakers; any other language value uses the English ones.
Fetch the live list at runtime instead of hardcoding:
clone_… ids, created in the dashboard) are not slots: a clone is one speaker in every language, and only API keys on the account that created it can use it.
Languages
Versa speaks nine languages.
Every voice slot works in every one of these languages (see Voices for which speaker a slot resolves to).
Fetch the live list at runtime rather than hardcoding it:
Which languages are understood
These nine are the languages a voice can speak, and speech recognition, which voice agents and the speech-to-text API use to understand a caller, covers all nine with one model: English, Spanish, Catalan, Galician, Basque, Portuguese, French, Italian and German. Recognition is told the agent’s language on every call, which keeps one-word replies (“sí”, “no”) reliable; an agent set to detect the language lets the model work it out.Models
versa-1.0 has been retired and is no longer served.
Parameters
The current model,versa-2.0, is a distilled model that runs a fixed schedule of 4 denoising steps without guidance on every request. Several request parameters are therefore accepted for compatibility but do not change the audio:
Speed
The model takes its timing from its own duration model, and there is no rate control in the current graph. Accepted for compatibility; no effect on versa-2.0.Steps
steps (called nfe server-side) is the number of denoising steps. The current model was trained for exactly 4 and always runs 4, on voice agents, WebSocket streams and REST alike. Accepted for compatibility; no effect on versa-2.0. On a live voice-agent session a per-session nfe can only be lowered, never raised above the default.
Voice character
cfg and cfg_text set guidance strength. The current model is trained to run without guidance: the voice comes from the reference speaker alone. Accepted for compatibility; no effect on versa-2.0.