Measured, Not Estimated

The Numbers Behind the Chart

Every score on this page comes from running the transcription engine on songs it was never trained on, then grading it against hand-checked references. Nothing here is a hand-picked demo — and every number is explained in plain English.

No account needed · Free · Runs in your browser

The Short Version

Four numbers that describe a whole transcription, from the first hit it hears to the last note of the chart — and what each one actually means for your song.

84% Hit Detection Score
69% Songs At Exact Tempo
94% Notes On The Beat
8 Parts Detected

The 84% is the average across the five voices that carry a groove — kick, snare, hi-hat, crash and ride — measured on songs the engine had never heard before.

What a score out of 100 actually means

Think of each figure below as a grade for one drum part. The engine is judged on every real hit in the reference recording, and it can fail in two ways: it can miss a hit, or it can invent one that is not there. Both mistakes cost it the same, which means it cannot inflate its grade by guessing extra notes to cover the gaps.

A score of 100 would mean it found every real hit and added nothing that was not played. A score of 91 means that for every hundred real hits, the misses and the false hits add up to nine. That is the number drummers care about: how much of your part shows up on screen, and how much is junk.

If you know the field: that balanced score is called F1. It is the standard way to report detection results in machine learning precisely because a single “accuracy” figure can be gamed by guessing, and F1 cannot.

Note Detection

How often the engine finds a real drum hit and files it under the right piece of the kit. Each instrument has its own model, trained on that instrument's sound — which is why a kick is near-perfect and the three toms are the hardest.

A hit counts as correct when it lands within a tenth of a second (± 100 ms) of the real one — tight enough that a note can never look visibly late on the highway. Kick and snare, the backbone of every groove, are found most reliably; hi-hat and cymbals follow. The three toms are the toughest part of any kit to tell apart, so a tom hit is usually found but sometimes filed under the wrong tom — the note is there, the label is what slips.

Tempo Detection

A chart is only playable if its beat grid lines up with the music, so tempo is judged separately from note detection. The detector sees only when notes fall — never the answer it is asked to produce.

Percentages are the share of songs where the detector got the answer exactly right. Two frozen held-out song sets, never used for any tuning decision. “16th grid” is the grid a chart actually draws, with every beat line fixed from the start of the song rather than nudged to fit.
Metric Validation Test
Songs at the exact tempo 69% 68%
Songs at the right tempo, allowing half/double time 82% 85%
Notes within 25 ms of the beat grid 94% 94%
Sections at the exact tempo 68% 71%
Typical distance from the grid 10.5 ms 10.3 ms

The two sets above are separate batches of charted songs, each with a known authored tempo, so you can see the figures hold up on data the engine never touched. One limit deserves to be stated plainly: about one song in six comes back at half or double time. That is not a bug being passed off as a feature — a beat heard in sixteenths and the same beat heard in eighths produce identical hits, so nothing in the note timing can decide between them. Given the right half or double, the engine nails the tempo 84% of the time. Rather than hide that behind a flattering number, we publish it.

Difficulty Generation

One transcription becomes four charts. The easier versions are written by a model that decides, hit by hit, whether a note belongs at that difficulty — and it is graded against human-made charts of the same songs.

Easy 85%

Keeps about 38% of Expert's notes.
Downbeat and backbeat only — the skeleton of the groove.

Medium 86%

Keeps about 55% of Expert's notes.
The groove plus simple fills, with the busiest passages thinned.

Hard 85%

Keeps about 79% of Expert's notes.
Near note-for-note — only the densest ornamentation comes out.

Expert Source

The transcription itself — every hit found, at the time it was played.
Graded by the note-detection scores above.

Each percentage is the model's grade against human-written charts for the same song, on songs held back from training — the model's validation split. The retention figures are how much of Expert each level keeps. Generation is deliberately note-preserving: an easier chart removes hits, it never invents them — so every note you play was really in the recording.

How These Numbers Are Produced

The method matters as much as the result, so here is exactly what is and is not being claimed.

Want the engineering detail? Read how the detection pipeline works → or start with the beginner's guide to drum transcription →

Hear It For Yourself

Numbers only go so far. Drop in a song you know by heart and see what the engine writes for it.