Voice agent teams face a persistent challenge: the fixed catalog of 400 voices rarely matches the specific persona a client demands—a Quebecoise receptionist for a Montreal dealership, or a narrator in his sixties with lecture hall gravitas. Briefs always outnumber any catalog, and traditional cloning fills the gap one speaker at a time, each requiring sourcing, consent, and licensing.
Gradium, the Paris-based voice AI company spun out of the Kyutai research lab, offers a different solution. Voice Design reads a written description and generates complete new voices in seconds—no reference audio, no speaker, no rights to clear.
Deployment ready: Voice Design is live in the Gradium API and in Studio, available free on all plans including the free tier. Voices you create run on the same streaming text-to-speech endpoint as catalog voices, with identical latency and output formats.
The Casting Brief as API
The description is the only input the model uses. Gradium's documentation lists the attributes it responds to—similar to a casting call: gender, age band, accent or origin, pitch, pace, energy, timbre, resonance, register, manner, and the role the voice serves. Descriptions range from 1 to 500 characters in English, French, Spanish, Portuguese, or German. Gradium recommends ending with the intended use, as it steers delivery and register rather than just the voice's tone.
A single request returns 1 to 5 candidates, typically ready within 3 to 5 seconds. Each candidate is a variation on one character; a different character requires a new description, not more samples.
From Candidate to Production Voice
The workflow involves four API calls. First, POST /voice-generator/generate creates candidate IDs with ready: false. Then, GET /voice-generator/embeddings polls until the process completes. Once ready, a second POST call selects the chosen candidate, and the final step integrates the voice with the existing text-to-speech endpoint, making it immediately deployable in real-time applications.
This streamlined path—from natural language prompt to production-ready voice—eliminates the friction of traditional voice acquisition, enabling businesses to prototype and launch custom voice agents in minutes rather than weeks. As of 2026, this approach aligns with broader industry trends toward generative AI for personalized user experiences, where flexibility and speed are critical for brand differentiation.
In practice, a team could specify: "A warm, authoritative male narrator in his sixties, with a British English accent, moderate pace, and calm energy, suitable for an audiobook," and receive a set of voices ready for immediate integration. This capability not only accelerates development cycles but also ensures that every voice interaction can feel bespoke, enhancing user engagement and brand consistency across automated touchpoints.
via MarkTechPost