How accurate is the audio analysis?
DJ Coach can analyse your audio in your browser to suggest a BPM, a key and a beat grid, to score energy, and to find drops, breakdowns and a likely outro. This page shows how we tested it, how it did, where it falls short and what the product does about that. Treat every suggestion as a second opinion and confirm it by ear.
Figures from the validation of 28 October 2026, measured on analyser version 1.3.0 unless a set below names another. The product now runs version 1.4.0; each set says what that changed.
The short version
- On clean material with known answers it is strong: the right key on 98/113 synthetic files, and no wrong key at a confidence of 0.35 or higher.
- On real music it agrees less: its bar lines match a reference tool on 9/20 tracks and its key on 8/20.
- So it is built to doubt itself. When it isn’t sure it makes no suggestion, and every card tells you to confirm by ear.
What we tested
We checked the analysis against two sets of audio.
- Synthetic files, with exact answers
- We generated 122 files (81.7 minutes) in which the true tempo, key, bar lines and sections are known exactly. That shows whether the analysis is right when the answer is certain.
- Real Creative Commons tracks
- 20 tracks (75.6 minutes), mostly electronic, including house, techno, dubstep, drum and bass and jungle. They are free to download from incompetech.com and archive.org, so you can see what was measured and how. Real music has no answer key, so we compared the analysis with reference tools for beats, bars, tempo and key. Those tools are used only for testing and are not part of DJ Coach.
The real-music figures measure agreement with those reference tools, not accuracy against what a DJ would mark. Where the analysis and a reference tool disagree, either one can be wrong.
The real set is small and lopsided: 14 of the 20 tracks are by one composer, only 2 are drum and bass or jungle, and only 1 is dubstep. Read the real-music figures as indicative.
On synthetic files with known answers
Measured on analyser version 1.1.0. Version 1.3.0 changed only 3 of the 122 synthetic files, each by a fraction of a BPM. Version 1.4.0 changed none of these figures, only the tempo and grid confidence of 4 files.
- Beats
- F-measure 0.909 on 87 files with known beats. The score rises as more true beats are found and fewer are invented.
- Tempo
- Within 4% of the true tempo on 86.2% of files, and on 100% when half, double, triple or a third of the true tempo also counts.
- Tempo on syncopated drums
- Two-step, half-time, breaks and hip-hop (28 files): 28/28 when half, double, triple or a third of the true tempo also counts. Two-step drum and bass and half-time are still read at half tempo, as described below.
- Bar lines
- The first beat of the bar was found on 83/87 files. F-measure for bar lines: 0.901.
- Key
- 98/113 exactly right, with no wrong key at a confidence of 0.35 or higher. Of the 15 misses, 7 were the relative major, 3 the parallel major, 1 a fifth below and 4 got no key.
- Sections
- Section changes (intro, build-up, drop, outro) found within a bar of the truth: recall 0.93, on only 5 files.
- Loudness
- Within 0.1 LU of a reference meter on 121/121 files. The other file is digital silence, which has no loudness to compare.
On real music: agreement, not accuracy
Agreement with reference tools on 20 real tracks, analyser version 1.3.0. Version 1.4.0 was not re-run on them, because the audio is not kept with the product. It changes results only for audio with no tempo, a beat tracked over fewer than 32 beats, or nothing above −70 LUFS, and each of these tracks has a tempo, runs at least 145 seconds and is far louder than that, so we expect the same figures. The reference here is another program’s output, not a person’s label.
- Beats
- F-measure 0.859 against the beat reference.
- Tempo
- Agreed with the reference tempo to within 4% on 75% of tracks, and on 80% when half, double, triple or a third also counts.
- Bar lines
- The first beat of the bar agrees on 9/20 tracks. F-measure for bar lines: 0.609.
- Key
- The same key as the key reference on 8/20 tracks, and 4 tracks get no key at all.
- Loudness
- Within 0.1 LU of the reference meter on 20/20 tracks.
Where it is weak
The known weaknesses, in plain words. Each one says what DJ Coach does about it.
- Cautious on real music
- On real tracks the analysis is often unsure of its own bar grid: 18 of 20 had a grid confidence below 0.5. Markers on such a grid aren’t snapped to bars, so most real-music markers stay low-confidence and few qualify for export. In DJ Coach: where your rekordbox or Traktor file already has a beatgrid at the matching tempo, markers are placed on that grid instead.
- Half-tempo drum and bass
- Two-step drum and bass and half-time tracks are read at half their tempo: 7/10 synthetic two-step files, 5/5 half-time files, and the real drum and bass track Mongrel. Their confidence is 0.45 or lower. In DJ Coach: a low-confidence tempo is never suggested, so you keep your file’s BPM.
- Tempo read at 2:3 of the reference value
- Some tracks come out at 2:3 of the reference tempo: Club Diver at 92.44 BPM against the reference’s 139.73, and Trips It, both at a confidence of 0.12 or lower. In DJ Coach: that is below the suggestion line, so no BPM is offered.
- Syncopated drums
- Hip-hop, drum and bass and other syncopated beats can fool the check for a steady percussive pulse. It happened on synthetic hip-hop and drum and bass, and on 5 real tracks whose grids agreed with the reference. When it does, tempo and grid confidence are capped at 0.45, so correct answers are under-trusted: of the 75 correct synthetic tempos, 20 sit at 0.45 or lower. In DJ Coach: such a track can show a correct tempo with low confidence, and low confidence means no suggestion.
- Bar lines on real music
- The first beat of the bar agrees with the reference on 9/20 tracks. On one track, Brain Dance, the analysis disagreed with the reference at a confidence of 0.73. In DJ Coach: “Confirm the phrase start by ear” stays on every card, and only cue points that sit on a downbeat at confidence 0.35 or higher are exported.
- Key on real music
- The key matched the key reference on 8/20 tracks, and 4 got no key. That reference is itself weak: on the synthetic files it was exactly right on 43/113 against 98/113 for DJ Coach, so a mismatch doesn’t prove the analysis is wrong. In DJ Coach: a key below 0.35 confidence is not offered as a suggestion.
- Track structure on real music
- Section changes were scored on only 5 synthetic files. On real music nothing was scored: a recent change moved section boundaries on 8 of the 20 tracks, with nothing to check them against. The confidence needed before the planner uses an analysed start point, 0.6, is a starting value, not an evaluated one. In DJ Coach: every card still tells you to confirm the start by ear.
- Energy on real music
- The energy score (rhythm, bass and brightness, not loudness) has not been compared with anything yet: not with DJs’ own ratings and not with a reference tool. In DJ Coach: a measured energy is used in the plan without an accept, except an uncertain one, which is not used; a rating you type in the table replaces it, and the card’s energy chip says “from your audio”.
- Vocals
- Vocals are not analysed at all. In DJ Coach: a plan with long blends says so once and tells you to listen for two vocals at once.
What the product does about it
These safeguards apply to every suggestion and every cue point.
- Low confidence means no suggestion. Below its confidence line, a BPM, key or bar start isn’t offered, and the values in your file stay as they are.
- “Confirm by ear” on every card. Each transition card tells you which bar to check by ear, even when the analysis is confident.
- Hollow markers. On the timeline, a marker with a confidence below 0.5 is drawn hollow and labelled “confirm by ear”.
- Only confident downbeat cues are exported. A cue point goes into your download only if it sits on a downbeat and its confidence is 0.35 or higher.
- Your corrections win. Anything you fix by hand stays fixed: the analysis never replaces it.
What is still unverified
These have not been checked yet, so treat them as unproven.
- Safari, iOS and Firefox. Whether files decode, how much memory a phone needs under the 200 MB and 20 minute limits for constrained devices, and how long the analysis takes have not been measured. The tests so far ran in Chromium-based browsers.
- Real rekordbox and Traktor import. We have not yet imported analysed cue points into the real programs to check that the “DJC …” memory cues land on the right beat, that your own cues are kept and that no hot-cue slot is touched. A decoder’s start offset can put a point tens of milliseconds off the beat your software shows, and the offset differs between MP3 and AAC/M4A.
- Real music labelled by people. The confidence levels were fitted on synthetic material. Re-fitting them needs a licensed set of real tracks that people have labelled, and more drum and bass and breakbeat tracks before the half-tempo rule changes.
- A by-ear check of real plans. Analysed start points at the estimated outro, entry cues and the energy curve have not yet been compared with a DJ’s own sense of the track.
Use it with your DJ software
DJ Coach is built to sit next to rekordbox and Traktor, not to replace the analysis you already trust.
- It starts from what your software found
- From a rekordbox XML or Traktor NML file, DJ Coach reads each track’s BPM, key, beatgrid and cue points, and plans from them.
- The analysis is optional
- Without it, the plan is computed from your playlist’s own data. With it, the analysis can suggest a BPM, key or beatgrid for a track your file has no value for, and it shows where it disagrees with the value your software found. Your file’s BPM, key and beatgrid never change until you accept a suggestion, and a value you corrected or locked is never replaced.
- Your grid places the cues
- Analysed cue points are placed on your file’s own beatgrid whenever the analysed tempo agrees with it (the same tempo, or half or double). A point within a quarter of a beat of one of your grid’s beats is moved onto that beat.
- Exports come back as memory cues
- Tick “Add cue points” when you download a rekordbox or Traktor playlist, and “Include analysed cue points” appears, if the analysis has cue points to offer. It is off until you tick it. The points are written as memory cues named “DJC …”, for example “DJC Drop”, and never as hot cues. Other formats can’t carry cue points, so they get none.
- Your own cues are kept
- Your cues travel in the same file, unchanged (an import replaces a track’s cues with the file’s, which is why they are there). A track whose cues can’t be carried over unchanged gets no analysed cue. Your music files are never changed.
- Not yet confirmed in the real programs
- We have not yet imported these files into rekordbox and Traktor to check all of this: see “What is still unverified”.
Think it got a track wrong? Tell us, or try it on your own playlist.