Gradium launches Voice Design: text prompts become synthetic voices in seconds

The Parisian startup Gradium, spun out of the Kyutai research lab, has released a capability that generates entirely new voices from a written prompt — no reference recording, no human speaker, no rights clearance required. The feature is live now in the company's API and studio, free on every plan including the free tier. Once saved, a voice runs on the same streaming TTS endpoint as any catalog voice, with identical latency and format support.

The description is the only input the model receives. Gradium's documentation lists the attributes it responds to: gender, age band, accent or origin, pitch, rate, energy, timbre and resonance, register and character, and the role the voice will play. Prompts can be 1 to 500 characters in English, French, Spanish, Portuguese or German; the company recommends ending with the intended use, because that steers delivery and register, not just vocal "color." A single request returns one to five candidates, typically within three to five seconds; they are variations on the same persona, so a different character requires a new description, not more samples.

Moving from candidate to production voice takes four API calls: POST to /voice-generator/generate creates candidate IDs with ready: false; GET to /voice-generator/embeddings polls until they flip; each candidate is then auditioned through the regular TTS endpoint using its ID as the voice_id; and POST to /voices/from-embedding promotes the chosen one. Candidates carry three restrictions that promoted voices do not: audition text is capped at 100 characters, they are accessible only via REST, and the WebSocket and Speech-to-Speech endpoints reject them. Unconverted candidates are deleted after 30 days. Promotion is free, removes the expiry, and consumes a custom-voice slot shared with clones — five on the free plan, 1,000 on paid plans. Generation is intentionally non-deterministic: the team expands the prompt first, and the expansion varies between requests, so the same prompt with a fixed seed still yields a different voice.

Gradium ran a blind pairwise listening test on accent prompts against five other Voice Design systems available via public API across five languages. Native speakers heard two unlabeled clips and chose the closer match, or declared a tie. Out of 7,627 comparisons, the company reports a 72.6% win rate for Gradium, 13.6 points above ElevenLabs' eleven_ttv_v3 at 59.0%, followed by Inworld at 44.8%, Fish Audio at 36.7% and MiniMax at 31.7%. Win rate is calculated as wins plus half the ties, so 50% is the baseline. Gradium placed first in all five languages. The widest gaps appeared in regional accents that most catalogs flatten: Quebec French 97%, Rioplatense Spanish 86%, Bavarian German 85%, Colombian Spanish and African Portuguese 83% each.

A model-based judge on the same prompt set agreed. Gemini 3.1 Pro scored individual unlabeled clips from 1 to 5 and produced the same ranking: Gradium 4.06, ElevenLabs 3.86, Inworld 3.64, Fish Audio 3.51. Separately, the product page claims 83.4% prompt adherence on the English split of InstructTTSEval, the academic benchmark for instruction following in TTS. The important caveat: all of these numbers were designed and run by the vendor itself.