Voice agent teams keep hitting the same wall. The catalog holds 400 voices and the brief asks for the one that is not in it: a Quebecoise receptionist for a Montreal dealership, a narrator in his sixties with lecture hall authority. Briefs outnumber any catalog, and cloning closes the gap one speaker at a time, each carrying sourcing, consent and a licence.
Gradium, the Paris-based voice AI company spun out of the Kyutai research lab, has shipped a different answer. Voice Design reads a written description and returns complete new voices in a few seconds. No reference audio, no speaker, no rights to clear.
Is it deployable? Yes, Voice Design is live in the Gradium API and in Studio, free on every plan including the free tier, and a kept voice runs on the same streaming Text-to-Speech endpoint as any catalog voice, at the same latency and output formats.
The casting brief is the API
The description is the only input the model gets. Gradium’s documentation lists the attributes it responds to, and they read like a casting call: gender, age band, accent or origin, pitch, pace, energy, timbre and resonance, register and manner, 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 intended use, because it steers delivery and register rather than only the colour of the voice.
One request returns 1 to 5 candidates, typically ready in 3 to 5 seconds. They are variations on a single character, so a different character means a different description, not more samples.
From candidate to production voice
The flow is four calls. POST /voice-generator/generate mints candidate ids with ready: false. GET /voice-generator/embeddings polls until they flip. Each candidate auditions through the ordinary TTS endpoint, using the candidate id as voice_id. POST /voices/from-embedding promotes the one you keep.
Candidates carry three restrictions converted voices do not: audition text is capped at 100 characters, they are REST only, and the TTS WebSocket and Speech-to-Speech reject them. Unconverted candidates are deleted after 30 days. Converting is free, clears the expiry, and uses one custom voice slot shared with clones. The free tier holds 5, paid plans 1,000.
Sampling is deliberately non-deterministic. Gradium team expands the description first, and that expansion varies per request, so the same prompt with a fixed seed still yields a different voice.
The benchmark, and how to read it
Gradium ran a blind pairwise listening test on accent prompts across six voice design systems reachable through public APIs and five languages. Native speakers heard two unlabelled clips and picked the closer match, or a tie. Across 7,627 comparisons, Gradium reports a 72.6% win rate against the field, 13.6 points ahead of 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 wins plus half of ties, so 50% is par. Gradium placed first in all five languages. The widest margins came on regional accents that most catalogs flatten: Quebecois French at 97%, Rioplatense Spanish at 86%, Bavarian German at 85%, Colombian Spanish and African Portuguese at 83%.
A model judge over the same prompt set agreed. Gemini 3.1 Pro rated single unlabelled 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. (Note: All of these numbers are vendor designed and vendor run.)
Key Takeaways
- Voice Design turns a 500 character description into up to 5 new voices in seconds, no reference audio required.
- It is live and free on every Gradium plan, in the API and Studio, across 5 languages.
- Vendor run blind tests put it at a 72.6% win rate over 7,627 comparisons, first in all 5 languages.
- Kept voices become a normal
voice_idon REST, WebSocket and Speech-to-Speech. - Sampling is non-deterministic, so an unsaved candidate is gone for good after 30 days.
Check out the Technical details and Docs. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us
Michal Sutter is a data science professional with a Master of Science in Data Science from the University of Padova. With a solid foundation in statistical analysis, machine learning, and data engineering, Michal excels at transforming complex datasets into actionable insights.
