acapellamaker acapellas that land in Serato ready

The accuracy numbers, measured

Everyone in this category says "accurate". Here is the actual test: every claim below was checked against a professionally analysed reference library of 2,579 real dance records, not a hand-picked sample.

WhatResult
Key exactly right87.5%
Key right or the relative key89.2%
BPM within 0.5%95.6%
BPM correct when claimed98.2%
Beatgrid correct when claimed95.3%

Why a sample would have lied

The first version of this was validated on 120 tracks and scored 99.1%. Run across all 2,579, the same code scored 94.8% and failed its own gate. Nothing about the code had changed — the sample was simply too small and too kind. Every number above comes from the full corpus.

Key: the parallel major problem

The single biggest failure mode in automatic key detection on electronic music is confusing a key with its parallel major or minor — F minor read as F major. With a general-purpose profile it happens on roughly one record in five, and it is the reason many DJs quietly stop trusting the key column.

Nine different tonal profiles were swept over the same audio. The one tuned for electronic dance music got 87.5% exactly right against 79.2% for the general-purpose default, and dropped parallel-mode errors to 1.7%.

Tempo: label accuracy is not grid accuracy

A tempo good to half a percent is a perfectly good label and a useless beatgrid. At 128 BPM, being 0.4 out drifts more than a full beat across six minutes — right at the start, wrong by the end.

Two independent estimators are used, with different strengths: one reliably gets the octave right, the other the period. They are folded to a common octave and required to agree; that agreement predicts correctness far better than either estimator's own confidence score. The agreed value is then snapped to the integer grid that 94.4% of this corpus genuinely sits on. Beat positions supply only the phase — never the tempo, because a beat tracker sometimes lays its beats at a different subdivision, and fitting them faithfully reproduces that error.

Three claims, three separate gates

Each claim stands on its own evidence rather than one all-or-nothing verdict:

  1. Key — claimed whenever there is one.
  2. Tempo — claimed only when the two estimators agree.
  3. Beatgrid — claimed only when the beats also land on that tempo, measured as a phase residual.

So a file can arrive keyed but tempo-less, or keyed and tempo'd but ungridded. That is the system working. The alternative — guessing to fill every column — is what makes DJs distrust the columns that are right.

Try it on one of your own tracks

Drop a track and hear the first 30 seconds free. No account, no card.

Open the tool

Questions

How accurate is the key detection?

Across 2,579 tracks cross-checked against a professionally analysed reference library, the key was exactly right 87.5% of the time, and 89.2% counting relative keys, which mix interchangeably.

How accurate is the BPM?

Within half a percent on 95.6% of tracks. When a tempo is actually claimed rather than declined, it is right 98.2% of the time.

What happens when it is not sure?

It says nothing rather than guessing. Key is claimed whenever there is one, tempo only when two independent estimators agree, and a beatgrid only when the detected beats also land cleanly on that tempo.

Related