Intelligibility on unseen text Two-ASR consensus semantic WER · Qwen3-ASR + Nemotron 3.5 · 400 matched prompts 0% 1% 2% 3% 4% 5% Supertonic 3 · M2 · 8-step 2.81% KittenTTS Nano · two-voice mean 3.07% Inflect-Micro-v2 3.99% Piper Low · two-voice mean 4.14% Inflect-Nano-v2 4.21% Supertonic 3 · M2 · 3-step 4.54% Lower is better. Whisper is excluded from this headline for every system; the complete three-ASR audit is reported separately.