MIDI Scale Finder

Find the most likely key/scale (currently Major/Minor only) from either:

• a dropped/uploaded .mid / .midi file, or
• a set of notes you pick manually (click the piano), including a computer-keyboard “piano mode”.

Live site: https://midi.hakono.me/

What the website does

• You can upload a .mid / .midi file, or build a note set by selecting notes on the piano (mouse/touch/keyboard).
• It reduces notes to pitch classes (C…B) and estimates the most likely tonic + mode.
• It shows a ranked list of candidate Major/Minor keys, with a best guess highlighted.

This is meant for quick “what key is this in?” checks for melodies, chord progressions, or any small set of notes.

How it works

Everything runs client-side.

1. Parse MIDI using the Tone.js MIDI parser (vendored as Midi.js).
2. Build noteWeights: pitchClass -> weight.
   • For MIDI upload: weight is total note duration across the whole file.
   • For manual note selection: weights are neutral (all selected notes count equally).
3. Score all 24 candidates (12 roots × Major/Minor):
   • add points for notes that are inside the scale
   • boost “important” degrees (tonic, dominant, subdominant, the mode-defining third)
     • the boost values are tuned with tune.js to maximize accuracy on a small(!) labeled set of 640 chord progressions.
   • subtract a penalty for notes outside the scale (and an extra penalty for the “wrong” third)
4. Sort by score and display the top result.

The “match %” shown in the UI means “how much of the input is inside the key” (by duration for MIDI, or by count for manual selection). It’s a descriptive metric, not a calibrated probability.

Limitations

• Major/Minor only. No modes, harmonic/melodic minor, blues scales, etc.
• Relative major/minor can flip (same pitch set). The weighted scoring tries to pick a tonic.
• Key changes/modulation: it assumes one key for the whole file.
• MIDI with lots of chromatic passing tones, borrowed chords, or dense percussion can confuse it (it currently doesn’t ignore drums).

On a small labeled set (older internal test run), it reached about ~88% top-1 accuracy. More testing should be done.

Contributing

This is a static site.

• Main UI: index.html, script.js
• Detection logic: scaleDetector.js
• MIDI parser: Midi.js (from https://unpkg.com/@tonejs/midi@2.0.28/build/Midi.js)

Run locally

Because the app uses ES modules (<script type="module">), you’ll usually want a local server (not file://). For example:

python -m http.server

Then open http://localhost:8000/.

Test / tune the scoring

• test.html is an internal evaluator (and optional in-browser grid search tuner).
  • Filenames must contain key + mode, e.g. C_Major.mid or G#Minor_MyChords.mid.
• tune.js is a Node script used to grid-search the scoring multipliers offline.

TODO

• Allow connecting external MIDI keyboards as an input source (support Web MIDI API).