While voice cloning has long relied on existing audio samples—as seen with Qwen3-TTS in llama.cpp—Gradium is shifting the paradigm toward generative synthesis. The Paris-based company, a spin-off from the Kyutai research lab, has released Voice Design, a system that generates entirely new synthetic voices based solely on written descriptions, eliminating the need for reference audio or licensing agreements with human speakers.

Add AlexTech.ai asPreferred Source on Google

From Casting Brief to API

Voice Design operates by translating a text prompt (up to 500 characters) into vocal characteristics. The model responds to specific attributes such as gender, age band, accent or origin, pitch, pace, energy, timbre, resonance, and register. Gradium suggests including the intended use of the voice in the prompt to better steer the delivery and register.

A single request generates between 1 and 5 candidates within 3 to 5 seconds. These are variations of a single character; creating a different persona requires a new description rather than more samples. The sampling process is non-deterministic: even with a fixed seed, the model's internal expansion of the prompt varies, meaning unsaved candidates are lost after 30 days.

Production Pipeline and Integration

The transition from a candidate to a production-ready voice follows a four-step API flow:

  1. POST /voice-generator/generate creates candidate IDs.
  2. GET /voice-generator/embeddings polls for readiness.
  3. Candidates are auditioned via the standard TTS endpoint using the candidate ID as the voice_id.
  4. POST /voices/from-embedding promotes a selected candidate to a permanent voice.

Once converted, these voices function as standard voice_id entities across REST, WebSocket, and Speech-to-Speech endpoints. The service is available in English, French, Spanish, Portuguese, and German. In terms of accessibility, Voice Design is integrated into the Gradium API and Studio, available on all plans including a free tier that allows up to 5 custom voice slots.

Benchmarking Regional Accents

Gradium's internal blind pairwise listening tests across 7,627 comparisons show a 72.6% win rate against other public API systems, including ElevenLabs (59.0%) and Inworld (44.8%). The most significant performance gains were observed in regional accents that are typically flattened by standard catalogs: Quebecois French (97%), Rioplatense Spanish (86%), and Bavarian German (85%). These results were mirrored by a model judge using Gemini 3.1 Pro, which ranked Gradium first with a score of 4.06.