Researchers have built a system that turns written descriptions of a speaker into the statistical building blocks voice synthesis needs.
The system, called ProPS (Prompted Profile Synthesis), takes a natural language prompt — say, "a thirties male speaker with an Indian accent" — and generates a distribution of speaker embeddings rather than a single fixed voice. It does this by converting the text description into a sentence embedding, then feeding it through a mixture density network trained to predict a Gaussian mixture model in the x-vector space. The model was trained to maximize the likelihood that its synthetic embeddings match real speakers who fit the described profile. Evaluations measured how well generated embeddings preserved requested attributes like age, gender, accent, and prosodic characteristics.
Most speaker embedding tools work in one direction: feed in audio, get out a vector. ProPS reverses that flow, which matters because TTS and voice conversion systems today typically require real voice samples to condition on. A text-driven approach could reduce that dependency and make it easier to specify speaker characteristics without sourcing recordings. Whether the outputs are distinguishable from real voices — and what that implies for misuse — the paper does not address.
The research sits alongside a growing push to make synthetic voice more controllable, but it is an academic framework, not a shipping product, and the gap between controlled benchmarks and robust real-world performance tends to be wide.