Hear every model on the same prompts — default voice and voice cloning, side by side. One clip per model, sourced from the highest-fidelity rig. Quality and prosody are obvious in seconds.
Time-to-first-audio, real-time factor, and memory for every model — per rig (windows-5090 · linux-3090 · mac-m4), cold and warm. Sortable.
Objective metrics for every model — UTMOS naturalness, WER intelligibility, and cloning-fidelity SIM. Sortable, default voice and cloning. Human votes remain the ground truth; these are backstops.
What each model can actually do — voice cloning, cross-lingual cloning, languages, expressive control, sample rate, license, CPU support. Toggle filters that combine, plus sort by any column.
Hear two clips blind and pick the better one — default voice or voice cloning. No install, ~5 seconds a vote, feeding a live human-preference Elo leaderboard. This is the ground truth the Scores lens is measured against.
