Mix Compare

/mix-compare — diff two versions of a mix, objectively

When you confirm a mix and a new bounce comes in (v4 → v5), this auto-aligns the two versions and diffs them per stem, per EQ band, in stereo image, phase, and noise — then translates the difference into a sentence you can hand straight back to the engineer.

It's git diff, but for audio. No single tool does this combo (time-align + stem diff + phase + plain-language explanation): reference A/B meters compare to a reference track, single-file "AI feedback" tools look at one file, aligners only align.

vs. single-file analysis

The point: a single-file tool answers "how is this mix?"; mix-compare answers "what changed in this revision, and why does it sound that way?"

Setup

pip install numpy scipy librosa soundfile   # required
pip install pyloudnorm                       # optional: LUFS / true-peak
pip install demucs                           # optional: per-stem diff (heavier, CPU)

Required deps give you alignment + phase diff + EQ + phase + noise. pyloudnorm adds loudness; demucs adds the per-stem breakdown. Without demucs, pass --skip-stems.

Trigger

1. /mix-compare <old> <new> — previous version first, current version second.
2. Two audio files + "compare the mixes", "what changed in v5", "what did they fix".
3. A revision arrives and you want it diffed against the prior bounce.

If two paths aren't given, ask for them in one line. Order matters: first = old, second = new (current). Every delta is new - old.

Workflow

Phase 1 — Run the compare

python3 ~/.claude/skills/mix-compare/compare.py "<old>" "<new>" --pretty --out /tmp/mixcompare.json

Options:

• --skip-stems — skip stem separation (align + diff + full-mix + loudness + phase + noise, ~20s). Quick check, no demucs needed.
• --diff-audio — write phase-invert diff wavs (full mix + vocal stem). Play them and you hear only what changed.

With stems, demucs splits both files (~3–6 min on CPU). Drop --skip-stems only when you need the per-stem breakdown.

Phase 2 — Read the JSON (delta-first)

Accuracy depends on the inputs. On two bounces of the same mix through the same chain (the normal v4→v5 case), alignment + phase-invert isolate the real change well. On two differently-mastered files, non-linear bus processing (comp/limit/saturation) breaks clean cancellation — phase_diff and fine per-stem deltas get noisy; the full-mix EQ delta and loudness stay reliable. Say which case you're in before drawing strong conclusions.

Interpretation heuristics (rules of thumb, not lab constants):

• per_stem.vocals.ms_width.delta > +0.3 dB → vocal got wider — a common physical cause of a vocal feeling "detached / floating" (often a room-reverb side effect).
• reverb_tail.delta negative → reverb was added but it got drier anyway (e.g. backing vocals pulled down removed more space than the reverb added).
• lowend_phase.*.phase_corr < 0.95 → low-end phase instability — a frequent cause of a mix that balances differently on different systems. Bass stem is usually ~0.99 (stable); if the other stem sits near 0.3, suspect pads / stereo reverb returns / layered synths with energy below ~120Hz → treat the low end as mono.
• high_noise.*.flatness > 0.4 → genuinely noisy. ~0.001 means no noise (so a "harsh highs" worry is unfounded — read it as a question to rule out, not a problem to fix).
• fullmix_octave_delta → which bands were touched. Presence (1.3–2.6k) / high-mid (2.6–5k) coming down reads as "smoother / softer".

Phase 3 — Translate to engineer-ready feedback

Turn numbers into language an engineer can act on:

1. One-line summary — the single biggest change in this revision.
2. Vocal change — level / width / space / EQ.
3. Problem + prescription (e.g. "other stem phase corr 0.27 below 120Hz → mono the low end on pads / reverb returns").
4. One follow-up question.

Don't

• State a structure label (intro / drop / …) as fact — that's a dance-music-biased heuristic, not a measurement.
• Dump raw numbers ("M-S +0.49, corr 0.27, flatness 0.0003") — translate to musical language.
• Lead with true peak / loudness. Engineers often deliver intentionally limited, loud bounces so a non-technical client hears it "finished." Only raise loudness if the user brings it up first ("why does this sound quiet?"). At the confirm stage, a limited loudness is normal practice, not a defect.
• Indiscriminately praise — give critique when asked, with the numbers behind it.
• Confuse file order — first = old, second = new, delta = new - old.
• Silently fabricate a fallback when a stage fails — say it failed and proceed with the rest.

Limits

• Stem bleed: htdemucs isn't perfect separation — vocal tails leak into other. Trust the old↔new delta over a stem's absolute number.
• Phase-invert diff: if the two versions differ in gain, the gain delta shows up in the diff. ratio_percent is an estimate of total change, not a clean isolation.
• Low-end phase corr is measured after stem separation, so separation artifacts can influence it — cross-check against the full-mix lowend_phase.
• demucs on CPU: two files = double the time. For a quick pass, --skip-stems.