Harmonic mixing and key detection: the Camelot wheel, and where the software gets keys wrong

Harmonic mixing means choosing the next track by its musical key, so that two records playing at once don't clash. It's a useful habit and an overrated rule. This guide explains the Camelot wheel, how key-detection software actually decides a key, and the predictable ways it gets it wrong. Sources were read on 30 Oct 2026.

1. The theory underneath: the circle of fifths

The circle of fifths arranges the twelve pitch classes in a cycle of perfect fifths. Neighbouring keys differ by just one sharp or flat, and every major key shares its key signature with a relative minor: C major and A minor both have none. Keys that share most of their notes are the ones that sound comfortable laid over each other. That is the whole idea behind harmonic mixing.

2. The Camelot wheel: the same circle, renumbered

The Camelot wheel replaces note names with a number from 1 to 12 and a letter. On Mixed In Key's chart, the inner A ring holds minor keys and the outer B ring holds major keys. 8A is A minor and 8B is C major; 9A is E minor and 9B is G major. Each step round the wheel is a step round the circle of fifths, and each A/B pair is a relative minor and major.

Mixed In Key's harmonic mixing guide lists the moves it considers safe:

• Same code, for example 8A to 8A.
• Same number, other letter, for example 8A to 8B (relative minor to major).
• One step on the same ring, for example 8A to 7A or 9A.

It also suggests walking round the wheel in one direction to keep a set "fresh and dynamic", and says bigger jumps such as 3A to 9A should be used "with extreme caution". Mixed In Key sells the software that popularised these codes, so read its guide as a vendor's method, not neutral research.

3. How key detection works

Most key detectors follow the same recipe, and the open-source Essentia library documents it plainly. Its KeyExtractor "computes HPCP frames for the input signal and applies key estimation". An HPCP (harmonic pitch class profile) is a 12-bin (or finer) chart of how much energy the audio has at each note name, folded across octaves. The Key algorithm then correlates that profile against a template for each of the 24 keys and picks the best match.

Two details matter:

• The template decides a lot. Essentia ships many profile types. Its docs describe edma as profiles extracted from electronic dance music, temperley as profiles from "euroclassical music", and krumhansl as profiles from listening experiments, a general option that works reasonably for pop. The KeyExtractor default is bgate.
• Confidence exists, but it's hidden. Essentia returns a strength value and a firstToSecondRelativeStrength: how clearly the best key beat the runner-up. Most DJ software shows you one confident-looking code and none of this.

4. Where it fails

These are the failure modes that follow from the method. Where a source states it, we cite it; the rest is our reading of how template matching behaves.

• Relative and neighbouring keys. A minor and C major use the same notes, so their profiles look alike. The top two candidates are often close, which is exactly what firstToSecondRelativeStrength measures. Expect 8A/8B and one-step confusions more than wild misses.
• Genre bias in the template. Essentia notes its edmm profile reports major keys as minor, because major modes are poorly represented in EDM. A template trained on one repertoire leans that way on everything.
• Tuning. The extractor assumes a reference (tuningFrequency, default 440 Hz) unless you correct it. A record mastered slightly sharp, or a vinyl rip played fast, can spill energy into the wrong bins.
• Music without a clear key. Drum tools, atonal intros, modal tracks (Dorian grooves, Phrygian techno) and songs that change key all get forced into one of 24 boxes. A single label for a whole track is an average.
• Pitch-shifting in the mix. If you use key lock off, or a tempo change shifts the pitch, the analysed key no longer matches what's playing.

5. What we think

Treat the Camelot code as a suggestion, not a verdict. It's most useful for spotting clashes in long, melodic blends, and nearly irrelevant for a 4-bar cut between drum-heavy tracks. When software says two tracks are compatible and your ears say they clash, believe your ears: the code is one number squeezed out of a correlation that may have been close.

Practical rules we'd follow:

1. Analyse your library with one tool and don't mix codes from different analysers; they disagree.
2. Spot-check the tracks you play most against a keyboard or a tuner app. Fix the label by hand when it's wrong.
3. Be suspicious of A/B and one-step "matches" that sound off; those are the tool's most likely errors.
4. Use energy and phrasing first, key second. A well-phrased mix in a "wrong" key often sounds better than a perfect key match on a clumsy transition.
5. Expect less from modal or key-changing tracks, and listen through the whole track before trusting its single code.

The gap we'd like a vendor to close: show the confidence score and the runner-up key. The data is there in the algorithm; hiding it makes a guess look like a fact.