Standard Voice warm RTF bench

Warm + measure on a fixed ~150-char English line. Gate: provider=webgpu and RTF ≤ 0.05 (~20×). Denoise dominates; production default is 5 steps (official WebGPU benches use 5).

Hello from Standard Voice on WebGPU. This warm timing line is about one hundred fifty characters long so we can measure realtime factor fairly today.

Status: Idle