Audio Render Forensics

Audio Render Forensics

Compare three renders without mistaking intended gain changes, latency, polarity, or sub-sample offset for distortion. Preserve the source files and write all analysis into a new output directory.

For a spatial pipeline, use $immersive-audio first to freeze channel/object/ scene representation, ordering and normalization, decoder or renderer, HRTF, and comparable output format. For mc-patcher, this skill may measure captured renders, but $mc-patcher still owns DAW→Bridge→Host routing, PDC/wet-return, loaded-binary identity, and hosted acceptance.

Required inputs

Obtain:

1. Original: the comparison-domain control, unprocessed by the target or candidate effect. Apply any required common domain transform first (for example, the frozen baseline decoder for a binaural comparison).
2. Target: the reference product/render to match or beat.
3. Current: the candidate algorithm/render under evaluation.

Require the same programme, sample rate, channel count, routing, automation, start position, and intended processing settings. If durations differ, identify whether the difference is silence/tail before interpreting metrics.

All three inputs must be in the same comparison domain. For AmbiX-to-binaural work, use three same-format binaural renders: a control decoded through the frozen baseline renderer, the target chain, and the current chain. Alternatively use a matching AmbiX-domain triplet. Never use raw multichannel AmbiX as Original against stereo binaural Target and Current renders.

Core workflow

1. Record absolute paths, bytes, sample format, SHA-256, sample rate, channels, frames, LUFS-I/LRA, sample peak and reconstructed true peak.
2. Estimate output delay from pre-whitened programme edges, detect polarity, and report the parabolic fractional-delay estimate. Inspect alignment_correlation.png before accepting any residual result.
3. Align to the Original grid. Default to integer alignment. Use --fractional-align sinc only when the fractional remainder is material and explicitly report that interpolation was applied.
4. Exclude outer filter/PDC boundaries and source regions below the configured silence threshold.
5. Compute both global scalar-fit residue and local common-gain residue at 100/50/20/5/1 ms. Treat the latter as a real-program nonlinearity proxy, not as pure-tone THD.
6. Infer the uncompressed gain plateau, GR distribution, GR event durations, 5 ms versus 50 ms gain modulation, 1 ms gain slew, attack/release movement, lobe-to-lobe gain jumps, and residue conditioned by GR depth.
7. Compute coherence-qualified transfer magnitude/phase, residual spectra, frequency-band residue, stereo S/M width, L/R gain mismatch, crest change, and pairwise Target-versus-Current correlation under raw, equal-LUFS, and equal-true-peak normalizations.
8. Generate worst-window CSVs and time-matched excerpts. Listen to those excerpts before recommending an algorithm change.
9. For literal THD/THD+N/IMD or LTI frequency-response claims, run controlled tones/sweeps using references/probe-suite.md.

Read references/methodology.md before changing any formula or interpreting a new class of processor. Read references/output-contract.md when auditing a run or assembling a report.

Run the analyzer

First choose a Python containing numpy, scipy, soundfile, and matplotlib. Also place ffmpeg on PATH for LUFS/LRA metadata. Run the example from the skill directory that contains this SKILL.md, so the relative scripts/ path resolves to the bundled analyzer.

python scripts/compare_audio_renders.py `
  --original "D:\audio\Original.wav" `
  --target "D:\audio\Reference.wav" `
  --current "D:\audio\Candidate.wav" `
  --target-label "Reference Limiter" `
  --current-label "Current Algorithm" `
  --output-dir "D:\analysis\candidate-vs-reference"

When the known transparent/static drive is +12 dB, pass it explicitly:

  --target-static-gain-db 12 --current-static-gain-db 12

When the processor should leave source peaks below a known boundary untouched, add --below-boundary-dbfs <value>; do not invent this boundary from the output.

Useful options:

• --true-peak-factor 32: independent high-resolution peak audit.
• --fractional-align report|sinc: report-only is the evidence-preserving default.
• --silence-threshold-dbfs -65: valid-programme gate.
• --edge-ms 100: outer-boundary exclusion.
• --fft-size 16384: Welch/CSD spectral resolution.
• --baseline-percentile 99.9: fallback GR baseline when static gain is unknown.
• --no-excerpts: omit copied listening excerpts when programme audio must not be duplicated.

Interpretation rules

• Never compare raw nulls before alignment and polarity checks.
• Never call a LUFS or peak mismatch “distortion.” Compare raw, equal-LUFS, and equal-TP results separately.
• Never call music-derived residue “THD.” It combines nonlinear waveform change, phase/SRC error, fast modulation, and noise after removing a local gain model.
• Never call a low-coherence music bin “frequency response.” Use only qualified bins, then confirm with tones or a sweep.
• Treat inferred GR as a model. Supply explicit static gain when known.
• Report all window sizes and denominators. Residue normalized to output is not interchangeable with residue normalized to source.
• Separate global mastering difference from local waveform damage.
• Preserve machine-readable JSON/CSV, plots, hashes, and excerpt manifests.
• State direct evidence, inference, and listening judgment separately.

Minimum delivery

Return:

• paths to REPORT.md and analysis_metrics.json;
• alignment delay/polarity/fractional remainder;
• LUFS-I and reconstructed TP for all three files;
• global plus 20/5/1 ms local residue for Target and Current;
• fast gain-modulation RMS, 1 ms gain-change p99, GR p99/max and >1 dB time;
• lobe jump p99, spectral-band residue, coherence-qualified transfer result, stereo width change, equal-LUFS correlation and RMS difference;
• the most diagnostic plot and listening-excerpt paths;
• a concise conclusion that respects the interpretation rules above.