Qaraami — the golden age, 1970s–1980s
When my teacher and I went looking for Somali sheet music, there was none. So I combined AI and my own music training so people and machines can finally read, play, and understand Somali music.
From a recording to a page
Press play. The engraving follows the audio note by note — the same binding the studio uses on full recordings.
Qaraami phrase in A pentatonic
Research demos
Short audio results from the project, each with what it is and how it was made. More are added as the research moves.
The melody was engraved as sheet music, parsed from its MIDI into 126 timed notes, and is performed here live in your browser by the platform’s own instruments: a plucked-string oud voice (a doubled course, two strings a few cents apart) over a durbaan frame-drum groove. Nothing is a recording — every note is synthesised the moment you press play.
Composed by Abdillahi Qarshe, 1950s · Sheet music by Khalid Ibrahim
A screen recording of the QaraamiGenAI studio — composing with the oud and durbaan instruments — will sit here.
The problem
Qaraami lived in voices and reels, never on paper. The tape hiss grows louder every year; the people who carry the melodies grow older. A tradition remembered by only a few is one silence away from ending — and what was never written down can still be lost.
cassette-era masters · decaying
The work
The pipeline separates the voice from the band, finds the pentatonic root, scale degrees, and beat grid, then engraves the melody as sheet music you can play, print, or export. Where the model is less certain, it prints lighter ink instead of pretending — the score never claims more than the recording gives it.
The signature
Every note carries the confidence of its own transcription: full ink at certainty, fading ink where the tape fades. Uncertainty is stated on the page, the way a careful edition would state it.
the ink is the confidence
The archive
A listening room with every source stated plainly. A studio where the score, the waveform, and the recording share one cursor. A research corpus behind both — across AI, computational ethnomusicology, and cultural preservation.
605
recordings surveyed in the research corpus
105
cassette tracks digitised and on hand
412
notes in the first end-to-end transcription
0.68
its mean confidence — printed, not hidden
Nothing here will ever be lost again.
End on a single held note