Gradium Launches Voice Design to Generate Custom Synthetic Voices From Text Prompts

The service answers a recurring problem for voice teams: a catalog of 400 voices still cannot supply the specific voice a brief demands, such as a Quebecoise receptionist for a Montreal dealership or a narrator in his sixties with lecture-hall authority. Cloning could fill such gaps, but each clone carries sourcing, consent and licensing overhead. Voice Design instead treats the written description as the only input.

According to Gradium's documentation, the model responds to attributes that 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 performing. Descriptions can range from 1 to 500 characters in English, French, Spanish, Portuguese or German. The company advises ending a description with the intended use, because that steers delivery and register rather than only voice quality. One request returns one to five candidates, typically in three to five seconds, and candidates are variations of a single character, so a different character requires a different description, not more samples.

Promoting a candidate to a production voice uses four API calls: generate candidate IDs with POST /voice-generator/generate; poll POST /voice-generator/embeddings until the IDs are ready; audition candidates through the ordinary TTS endpoint using the candidate ID as the voice ID; and promote the chosen voice with POST /voices/from-embedding. Until conversion, candidates carry three restrictions: audition text is capped at 100 characters, they are available only through REST, and both the TTS WebSocket and Speech-to-Speech endpoints reject them. Unconverted candidates are deleted after 30 days. Converting a candidate is free, removes the expiration, and uses one custom voice slot shared with cloning; the free tier holds five slots, while paid plans hold 1,000.

Sampling is deliberately non-deterministic. Gradium's team first expands the input description, and because that expansion varies per request, even the same prompt with a fixed seed will yield a different voice, according to the report.

Gradium evaluated Voice Design with a blind pairwise listening test on accent prompts against five other voice design systems reachable through public APIs, across five languages. The company said native speakers heard two unlabelled clips and picked the closer match or declared a tie. Across 7,627 comparisons, Gradium reported a 72.6 percent win rate, calculated as wins plus half of ties, where 50 percent is par. That put it ahead of ElevenLabs' eleven_ttv_v3 at 59.0 percent, Inworld at 44.8 percent, Fish Audio at 36.7 percent and MiniMax at 31.7 percent. Gradium placed first in all five languages. Its widest margins came on regional accents that most catalogs flatten: Quebecois French at 97 percent, Rioplatense Spanish at 86 percent, Bavarian German at 85 percent, and Colombian Spanish and African Portuguese at 83 percent. A model judge, Gemini 3.1 Pro, rated single unlabelled clips from one to five and produced the same ranking: Gradium 4.06, ElevenLabs 3.86, Inworld 3.64 and Fish Audio 3.51. Gradium's product page also claims an 83.4 percent prompt-adherence score on the English split of InstructTTSEval, an academic benchmark for instruction following. The report notes that all of these numbers are vendor-designed and vendor-run.