The shortlist Taona pulled from the v2 roster, re-rendered with best-of-32 sampling, denoised references, and per-intent prosody calibration. Studio renders and the 8 kHz narrowband telephony renders the dev will actually ship.
Honest read. Studio quality lifted +0.09 over v2, but the 8 kHz phone render regressed -0.17 (v2 3.12 → v3 2.95). The higher sampling buys studio gain that the narrowband codec eats. Mei is broken in v3 (studio 2.88, phone 1.90) despite the reference clip being identical to v2 — the v3 inference settings don't play nicely with that voice. Mark her as such and either re-render her with v2-style settings or drop her from the dev handoff. Aisha / Fatima are borderline on phone — listenable, soft.