Skip to content

0002 — Sample-exact filtering, no FFT

0002 — Sample-exact filtering, no FFT acceleration

Date: 2026-08-18 Status: Accepted

Context

The meter must reproduce the ITU-T G.191 reference implementation (sv-p56.c) within ±0.1 dB, pinned by conformance tests against reference fixtures. FFT-domain filtering or downsampled envelope analysis could speed up the pre-filter and envelope recursion, but would change the numerical result and break sample-exact conformance.

Decision

All filters run sample-serially in f64/f32 exactly as the reference does: the envelope is a cascade of two first-order recursions, the protection pre-filter is a transposed direct-form II biquad cascade with per-rate prewarped coefficients. No downsampling, no per-block power averaging, no FFT path.

Consequences

  • Benchmark RTF is ≈ 0.001–0.002 (≈ 500–1000× faster than realtime), so acceleration is pointless for this workload anyway.
  • Conformance is structural, not tuned: the same arithmetic as the reference, hence matching results by construction.
  • If future targets ever need SIMD, the exact sample-serial path stays the correctness baseline any accelerated path must match bit-for-bit.