A voice can be made from a short recording of a person and used like any other voice ("voice": "<id>").
The result is a similar-sounding voice, not an exact copy of the person. You must have the person’s consent.
A voice belongs to one language: record the person speaking that language, and the voice speaks only that language. It works best with 10-30 s of clean speech of one person, recorded with a microphone (not through a phone call), in a quiet room.
It is decoded, its silences are trimmed and it is levelled. The best 5-8 s become the voice. Before anything is stored, the new voice speaks three test sentences; if it fails them the request answers 422.The uploaded recording is not kept. Only the voice state and its details (name, language, gender, consent record, quality numbers) are stored. DELETE removes them for good. A voice belongs to the account of the API key; other accounts get 404.
Not 10-30 s long; less than 5 s of speech; too noisy (estimated SNR under 20 dB); telephone band (nothing above 4 kHz); the voice fails the test sentences. The message says why. Several speakers, music and strong echo are not checked: send one person’s speech only.
POST /v1/audio/speech with "voice": "<id>" or "voice": {"id": "<id>"}. The language is the voice’s; another language answers 400. An id that is not yours answers 404 and never falls back to a default voice. After a model update a voice may answer 409: create it again.
