BREAK BOSS

Hip-hop was built on breakbeats: the few seconds on an old record where the band drops out and the drummer plays alone. Producers dig through thousands of records to find them. Break Boss does the digging. Give it a song and it hands back the sections that work as breaks.

I built the first version in 2020 as my Master’s thesis at the University of Miami. In 2026 I rebuilt it twice: once the way it was, so I could test it properly, and once from scratch with current methods. The new one runs in your browser, below.

Drop songs here

MP3, WAV, AIFF, M4A, FLAC and most other audio files. Drop a stack of full songs. The breaks come out the other side, surest first.

Your audio never leaves your computer. The model runs on this page.

2020

The thesis version

Every clip is turned into a colour picture of its sound, a spectrogram. An image network that was trained on photographs (ResNet34) is then retrained to sort those pictures into breaks and non-breaks. I called it the Perceptron Audio Classifier.

To search a song, a second tool estimates the tempo, cuts the song into two-bar slices starting at 0:00 and classifies every slice. That was the Break Identification Genius.

  • Learns fromabout 900 hand-cut clips from 288 records
  • Size21 million weights, 85 MB
  • Speedabout 100 seconds a song in 2020
2026

The rebuild

A break is mostly a stretch where the drums are playing and little else is. A large separation network (Demucs) can pull a mix apart into drums, bass, vocals and the rest, so it can mark, moment by moment, how much of a recording is drums. On my hand-labelled clips that measure alone separates breaks from non-breaks almost perfectly, but the network is far too big for a web page.

So it acts as a teacher. It labels records from my own collection, and a small network learns from those labels, from my hand-cut clips and from whole songs where I had marked the break. That network scores a song every fifth of a second. The finder then picks out each drum stretch and the best loop of whole bars inside it.

  • Learns from835 hand-cut clips, 95 whole songs and 158 teacher-labelled excerpts
  • Size125,000 weights, half a megabyte
  • Speedabout a second a song, in your browser

Break or not? Short clips

Each version labels short clips as break or non-break. The honest test uses 58 records that were kept out of training entirely.

2020 thesis, as reportedsame records in training and test: 178 of 180 178/180
2020 version, rebuiltthe thesis's own split: 177 of 182 97.3%
2020 version, rebuiltrecords it never heard: 174 of 181 96.1%
2026 versionrecords it never heard: 178 of 181 98.3%
As reported in 2020 2020 version, rebuilt 2026 version

A model that hears only a clip's average tone, with no rhythm at all, scores 93.5% on unseen records. Telling a drum solo from a full band in a short clip is not the hard part.

Find the break in a whole song

The test a producer cares about. Each version searches 21 whole songs it never trained on, where I had marked the break by hand.

2020 version, rebuiltits top pick is a marked break: 16 of 21 76%
2026 versionits top pick is a marked break: 17 of 21 81%
2020 version, rebuilta marked break in its top three: 16 of 21 76%
2026 versiona marked break in its top three: 19 of 21 90%
As reported in 2020 2020 version, rebuilt 2026 version

A pick counts when it overlaps a hand-marked break by at least half. The 2020 version handed back 5% of a typical song and the 2026 version 9%. Neither returned a passage I had marked as not a break.

What the numbers do and do not say

The thesis reported just under 99%: 178 of 180 test clips, with clips from the same records on both sides of the test. Rebuilt and tested on records it had never heard, the same design scores 96.1%. The original number was a little generous, and the idea held up.

The 2026 version scores higher on both tests with about 170 times fewer weights, small enough to run on this page. The margins are small, though: four clips out of 181, and one song out of 21 for the top pick. On tests this size the two are close, and the clear gains are size, speed and the top-three result.

Both versions were trained and tested on soul, funk, rock and jazz records from one compilation of known breaks, so expect weaker results on other kinds of music. A break here means drums in the clear. Whether it is worth sampling is a matter of taste, which no test on this page measures.