Design for a tool that finds the EQ crossover frequency for a transition by measurement and a psychophysics harness, replacing the slowest and most frustrating loop in manual mix authoring.
Links: DJ EQ Blending Technique, Programmatic DJ Mixing Tools, CyborgDJ, Camelot From YouTube, Planner-LM Composites, The Three-Layer Method
Status: design, not yet implemented. No numbers here are measured — every quantity is a parameter to be determined by the harness this page describes.
The February 2026 implementation results identified crossover frequency as “the single biggest lever,” found 300 Hz leaked bass harmonics into the mid band, fixed one track pair at 450 Hz, and concluded it “needs per-transition tuning or spectral analysis.” Finding that number by ear was the slowest part of the loop — each candidate cost a render and a listen, and the comparison was between two things heard minutes apart.
The goal is not a better guess. It is to stop spending judgment on the number at all, so the conversation moves up an altitude — testing technique rather than arguing about a few Hz.
H1 — the crossover is derivable, not searchable. With source-separated stems, the question “where do I split so that killing the low band removes the bass without gutting the body” is answerable from two curves: the bass stem’s spectral envelope and the melodic stem’s. The crossover wants to sit above the bass stem’s harmonic reach and below the melodic content’s fundamental region — a crossing point read off a plot, not a value hunted by trial. The February finding is this in hindsight: 300 Hz sat below the bass’s harmonic reach, 450 Hz cleared it.
If H1 holds, the search collapses into a measurement, and the per-bar third-octave matrix from
spectral_analysis.py already supplies the raw material.
H2 — the perceptual tolerance is coarse. A working DJ sets this by twisting a knob, in a loud room, under headphone bleed, mid-performance. That practice is independent evidence about the tolerance: knobs overshoot constantly and nobody notices except a studio engineer. If the just-noticeable difference is ±100 Hz rather than ±10 Hz, then precision beyond that band is wasted effort, and much of the February frustration was chasing resolution below the noise floor of human judgment.
H1 and H2 are not in tension — they are two independent routes to the same outcome. H1 says compute it; H2 says you didn’t need it that precisely anyway. Either one shortens the loop; both together close it. H2 is also the cheaper to test, and it bounds how much H1 has to deliver.
One tool serves both hypotheses and produces the training data for a learned evaluator as a by-product. Every trial is a labelled pair — the calibration harness is the A/B log.
The naive design (fixed 100 Hz steps, sweep upward until it sounds right) is the psychophysics method of limits, which carries known biases:
| Failure mode | Effect | Mitigation |
|---|---|---|
| Hysteresis — ascending and descending sweeps disagree | habituation overshoots going up, anticipation undershoots coming down; a one-direction sweep bakes in a systematic offset | interleave both directions, average the thresholds |
| Fixed grid wastes trials | equal effort spent far from the threshold as near it | adaptive staircase: start at 100 Hz, halve the step on each judgment reversal; converges in ~10–15 trials |
| Expectation contamination | knowing which sample is 450 Hz means partly grading your own hypothesis | blind presentation, randomized order, harness records the mapping |
Insert repeat trials — the same pair, presented twice, separated in the sequence. The rate at which the same choice is made is the noise floor of the entire project.
If self-consistency collapses at ±50 Hz, then ±50 Hz is not a meaningful target and no amount of tooling will make it one. This single number sets the resolution worth optimizing to, converts “stop arguing about a few Hz” from an opinion into a measured bound, and is obtainable in one afternoon. It is the highest value-per-effort item in this design, and it should be run first — it may retire H1 before H1 is built.
Once a few hundred pairwise judgments exist, do not train a model yet. First check which already-computable deterministic metric best predicts the choices:
If one of these predicts picks at high accuracy, the critic is done — explainable, deterministic, and no model required. Only if none predicts well is a learned evaluator justified, and by then both the training data and the evidence that the simple version was insufficient are in hand.
This is the three-layer discipline applied to the metric itself: try the deterministic, lower-altitude check before the opinionated one, and let it gate the write.
Crossover frequency is a mixer constraint, not a law of mixing. It exists because hardware gives three knobs. In the stem-rendering path there is no crossover at all — the incoming bass stem is muted outright, so there is nothing to tune and no harmonic leak to fight.
This tool therefore matters for the EQ-domain render path, which remains live because stem re-summation artifacts are unresolved and melodic content separates poorly (the “other” stem caps at 8.7–9.0 dB SDR). Both paths stay open until measured.