Voice agents often stall when a project requires a specific persona that doesn’t exist in standard libraries. Think of a Quebecoise receptionist for a Montreal dealership or a narrator with the gravitas of a sixty-year-old professor. Gradium, a Paris-based AI firm spun out of the Kyutai research lab, claims to solve this bottleneck without needing reference audio or complex licensing deals. Their new tool, Voice Design, generates unique voices from written descriptions in seconds.
The process is straightforward but technically nuanced. Users input a description of up to 500 characters in English, French, Spanish, Portuguese, or German. The model considers attributes like gender, age, accent, pitch, and even the intended job role. Within three to five seconds, the system returns up to five candidates. These are variations on a single character, meaning you get different takes on the same prompt rather than entirely new personas. Once a user selects a favorite, they promote it to a permanent voice ID. This converted voice then works on all endpoints, including WebSocket and Speech-to-Speech, just like any catalog entry.
Gradium backs its launch with aggressive benchmarking. In a blind pairwise listening test involving 7,627 comparisons across five languages, native speakers preferred Gradium’s output 72.6% of the time against competitors. This placed them 13.6 points ahead of ElevenLabs’ eleven_ttv_v3 model, which scored 59.0%. The margins were particularly wide for regional accents that traditional catalogs often flatten, such as Quebecois French (97% win rate) and Rioplatense Spanish (86%). A separate evaluation by Gemini 3.1 Pro confirmed this ranking, giving Gradium a 4.06 score compared to ElevenLabs’ 3.86.
There are caveats to consider. The sampling is non-deterministic, so even with a fixed seed, the same prompt yields different results. If you don’t save a candidate within 30 days, it vanishes. Also, these benchmarks are vendor-run, which always warrants a grain of salt. Yet, the ability to generate compliant, high-quality voices without clearing rights or sourcing actors represents a significant shift. It moves voice synthesis from a library retrieval problem to a generative design challenge, potentially lowering the barrier for hyper-localized content creation.
