Public test set · Same songs for both · Measured 15 Sept 2026
We tested two BPM finders on 662 songs
The same 662 public test songs through AudioForges and Tunebat, scored against their true tempo. AudioForges got 75% exactly right, Tunebat 49%.
Exact BPM, 662 songs
AudioForges
75%
TempoCNN plus our tuning
Tunebat
49%
Free online analyzer
Results
What we measured
| Test | Songs | AudioForges | Tunebat |
|---|---|---|---|
| Exact BPM (within 3%) | GiantSteps Tempo, 662 tracks | 75% | 49% |
| Exact key | GiantSteps Key, 58-track sample | About 50% | About 50% |
Both tools were run on 15 September 2026. Most AudioForges misses are half or double the true tempo, which describes the same beat and is easy to fix by ear.
Method
How the test was run
We used GiantSteps Tempo, a public research set of 662 electronic dance music previews, each labelled with its true tempo. Every song went through both tools unchanged, and each reading was scored the same way: within 3% of the true tempo is exact, anything else is a miss.
AudioForges was run through the same pipeline the site uses today. Tunebat was run through its free online analyzer. We picked a public set so anyone can repeat the test, and we did not drop any song from it.
The key comparison used a 58-song sample of GiantSteps Key. Both tools landed at about 50%, so we call it level. Key detection on electronic music is hard for every tool, and the most common miss is the relative major or minor.
Under the hood
What runs inside AudioForges
Tempo comes from TempoCNN, a published neural network by Hendrik Schreiber and Meinard Müller, run through Essentia under a commercial licence. We did not train it. Our work is around it: preparing the audio, settling half and double readings, and three backup detectors for songs the model cannot read, such as ambient intros. Key is read by two detectors, and the result tells you when they disagree.
Read more in how key and BPM detection works, or try the key finder and BPM finder on your own songs.
Frequently asked questions
Of the two we tested on the same 662 public songs, AudioForges: 75% of tempos within 3% of the true value, against 49% for Tunebat. Other tools were not part of this run, so we make no claim about them.
A reading within 3% of the annotated tempo. For a 128 BPM song that is anything from about 124 to 132. A reading at half or double the tempo counts as a miss here, even though it describes the same beat.
Public research datasets of electronic dance music previews, each labelled with its true tempo or key. Researchers use them to compare tempo and key detectors, which is why we used them instead of picking our own songs.
No. Tempo comes from TempoCNN, a published neural network by Hendrik Schreiber and Meinard Müller, run through Essentia under a commercial licence. Our part is the pipeline around it: how the audio is prepared, how half and double readings are settled, and three backup detectors for songs the model cannot read.
The key comparison was run on 58 tracks, where both tools landed at about 50%. That is too small to call a winner, so we report it as level. A run on the full key set will replace it on this page.
Yes. Drop any song on the BPM finder, free and with no account, and compare the number with the tool you use now.