Compare · MedASR Bench

02.1 · Head-to-head

Two models, read closely

Pick a baseline and up to two challengers — all measured on ViMedCSS. Deltas are challenger minus baseline — green improves on the baseline, red regresses. Models without measurements stay pickable and show their evidence gap.

02.2 · Efficiency

Accuracy against throughput

RTFx — audio seconds per wall second on platform hardware — against WER on ViMedCSS, with bubble area scaling by parameter count. Track E (sub-1B, CPU, under 8 GB VRAM) is config-ready.

2 of 17 systems measured — RTFx only ships with platform-verified runs; the rest of this plane is the evidence gap, not the field.

RTFx = audio seconds ÷ wall seconds, measured on platform hardware. Bubble area scales with parameter count. ViMedCSS test split · normalizer v0.1.

The Pareto frontier

Two models, read closely

Accuracy against throughput

The Pareto frontier

Two models, read closely

Accuracy against throughput