Umi

Organizations developing voice AI models face a fragmented evaluation landscape. Training teams lack standardized benchmarks for measuring whether generated speech matches requested voice characteristics—gender, tone, emotion, pitch, and accent. Quality assessment remains subjective, reliant on in-house panels, and difficult to compare across competing model architectures. Deploying and evaluating models at scale requires infrastructure investment that smaller research teams cannot justify. The status quo creates inefficiency: good models go unrecognized because evaluation is localized and non-comparable.

A distributed evaluation network solves this by creating open, measurable standards. Instead of isolated evaluations, multiple independent assessors run identical quality checks on all submitted models. Each evaluator scores models on content accuracy, audio quality, and adherence to voice trait specifications. Results are aggregated across evaluators to identify which models perform best, creating a transparent, reproducible leaderboard. This approach removes central gatekeeping and surfaces genuine improvements in model performance.

Vocence implements this as a decentralized subnet where miners deploy voice models (starting with prompt-based text-to-speech) and validators execute a shared evaluation pipeline. The Bittensor network coordinates registration and distributes rewards to both high-performing miners and validators who run honest evaluations. All models, evaluation code, and results remain open and auditable.

Vocence is a Bittensor subnet for decentralized voice intelligence development, currently focused on Prompt-based Text-to-Speech (PromptTTS). Miners train and deploy voice models that generate speech matching natural-language prompts describing content and voice traits (gender, tone, emotion, pitch, etc.), while validators evaluate model quality across content correctness, audio quality, and prompt adherence using a standardized evaluation pipeline and stake-weighted global consensus scoring.