Gradium Launches Voice Design: Write a Prompt, Get a Brand New Synthetic Voice in Seconds
Voice agent groups maintain hitting the identical wall. The catalog holds 400 voices and the temporary asks for the one that’s not in it: a Quebecoise receptionist for a Montreal dealership, a narrator in his sixties with lecture corridor authority. Briefs outnumber any catalog, and cloning closes the hole one speaker at a time, every carrying sourcing, consent and a licence.
Gradium, the Paris-based voice AI firm spun out of the Kyutai analysis lab, has shipped a completely different reply. Voice Design reads a written description and returns full new voices in a few seconds. No reference audio, no speaker, no rights to clear.
Is it deployable? Yes, Voice Design is reside in the Gradium API and in Studio, free on each plan together with the free tier, and a stored voice runs on the identical streaming Text-to-Speech endpoint as any catalog voice, on the similar latency and output codecs.
The casting temporary is the API
The description is the one enter the mannequin will get. Gradium’s documentation lists the attributes it responds to, they usually learn like a casting name: gender, age band, accent or origin, pitch, tempo, power, timbre and resonance, register and method, and the job the voice is doing. Descriptions run 1 to 500 characters in English, French, Spanish, Portuguese or German. Gradium advises ending with the supposed use, as a result of it steers supply and register reasonably than solely the color of the voice.
One request returns 1 to five candidates, sometimes prepared in 3 to five seconds. They are variations on a single character, so a completely different character means a completely different description, no more samples.
From candidate to manufacturing voice
The stream is 4 calls. POST /voice-generator/generate mints candidate ids with prepared: false. GET /voice-generator/embeddings polls till they flip. Each candidate auditions by means of the unusual TTS endpoint, utilizing the candidate id as voice_id. POST /voices/from-embedding promotes the one you retain.
Candidates carry three restrictions transformed voices don’t: audition textual content is capped at 100 characters, they’re REST solely, and the TTS WebSocket and Speech-to-Speech reject them. Unconverted candidates are deleted after 30 days. Converting is free, clears the expiry, and makes use of one {custom} voice slot shared with clones. The free tier holds 5, paid plans 1,000.
Sampling is intentionally non-deterministic. Gradium workforce expands the outline first, and that enlargement varies per request, so the identical immediate with a fastened seed nonetheless yields a completely different voice.