SOUND WAVE EDITOR
GREEDY SINUSOIDAL DECOMPOSITION · MATCHING PURSUIT
⊕ LOAD AUDIO
⇑ LOAD .SWE
♪
no audio loaded
POS
0.00 / 0.00 s
PLAY RANGE
FROM
s
TO
s
SHIFT
▶ Matching Pursuit Analysis
THIS RUN
—
TOTAL
—
MATCH
—
PER OCTAVE
COMPONENTS
1
0
active: 0
captured: 0%
fewer components = a rougher song but a livelier after-song; more = the song almost exactly, and an after-song that settles into a wash. Drag this, then press PLAY.
FFT WINDOW
the song's own length decides the frequencies; the shorter options are only for a quick look
MODEL THE ENDING
asking the model to explain the silence after the song is what makes the after-song fade and swell instead of sitting still — it costs some accuracy on the song itself
PIER
PILLARS
AT LEAST
cents apart
OVER
sweeps
A LOUD PILLAR
each extra sweep halves the gap and starts over, so the pier goes up in courses of increasing density instead of one thin skeleton — and every course is its own widely spaced set. The pillars are still the song's own loudest frequencies — the only rule added is that no two may land within the same interval, and the gap is an interval rather than a number of hertz so it means the same thing at 80 Hz and at 8 kHz. Few, widely spaced waves come back into alignment every few seconds, so the pattern they make keeps returning after the song ends. The deck is the ordinary dense search that follows, and it is what makes the song itself accurate. GPU only.
SPLIT INTO
costs N× the time — the bands run one after another, and choosing edges imposes structure the song did not ask for. Leave at 1 unless experimenting.
ENGINE
detecting GPU…
FIND
components
and/or
FOR
seconds
and/or
UNTIL
% captured