A buddy of mine is running a marathon and asked me for a playlist. Five hours and twenty minutes later, the first track was in his hands: Andy Statman’s Flatbush Waltz — a sparse klezmer mandolin instrumental — turned into a sacred-and-profane EDM banger that literally paces a runner. I can’t read music. So I built it the way I build everything now: I described what I was hearing to an AI composer, it named the fundamental I was fumbling toward, and we fixed that. Here’s the theory behind every decision — worked out live between a guy with ears and no vocabulary, and a composer with all the vocabulary and no ears at all.

Don’t rebuild the pig — flip it

The obvious way to make an “EDM version” of a song is to transcribe it: pull out the melody, quantize the notes, rebuild a new track from synths that quote the original. I tried that first. It kept coming out dead. A transcription is a skeleton, and a skeleton isn’t recognizable — the ear keeps asking for the record it knows.

So I threw the skeleton away and made the opposite bet: keep Statman’s actual recording — his mandolin, his phrasing, his room — and flip it. Recognizability stops being a problem to solve, because it’s no longer manufactured. It’s the record itself, wearing a new body.

A flip reworks the original recording. A cover re-performs the composition with new hands. Statman’s actual performance is the foundation here, so this is a flip, not a cover — and keeping his performance is the whole trick.

You can’t bolt a metronome to a heartbeat

Here’s the problem that ate most of the afternoon. A four-on-the-floor kick is a metronome — perfectly even, forever. A human plays rubato — leaning into phrases, pulling back at cadences, speeding up and slowing down on purpose. Lay a rigid grid over an elastic performance and they drift apart within seconds. My feedback, verbatim, was: “the kick is stoned while the band’s on coke.”

The fix isn’t a better metronome — it’s to iron the rubato out of the source first. Beat-track the recording (find where its beats actually land, unevenly), then time-stretch each individual beat to the same length, flattening the whole performance to one constant tempo. Then the rigid kick locks, because now there’s a fixed pulse to lock to.

That per-beat stretch is a phase vocoder — it re-spaces audio in time while preserving pitch. It’s exactly what a DJ’s software does to beatmatch two records; I just pointed it inward and made the record agree with itself.

Then tempo becomes a color

Once tempo is something you set instead of something you inherit, it’s a compositional parameter like any other. So I set it per section. The intro doesn’t start at the body tempo — it starts at 80 BPM and accelerates to 120 across the first ninety seconds, a long slow accelerando that, on a running track, reads as a body spinning up from a walk. The drops jump to 160. The outro decelerates. The tempo curve isn’t housekeeping; it’s the program. It’s my buddy’s legs over the length of a run.

The drop — and its opposite

EDM form is one idea run with discipline: tension and release. You withhold the full texture, build pressure, and at the drop you deliver everything at once. This track has two, placed by ear at the tune’s own structural seams. Between them sits a breakdown — elements subtracted to reopen the tension so the second drop can land again.

The subtlest move is at the very end, and it’s also subtraction: instead of letting the kick fatigue out under the finish, I cut it cold, eighteen and a half seconds from the end. Removing the beat is a bigger event than any addition could be — the ear snaps to what’s left, and the track exhales. Density is a dynamic just like loudness; the most dramatic thing you can do to a wall of sound is take part of it away.

Groove: the pump and the interlock

Two things give it motion. The first is sidechain compression — the whole mix ducks under each kick and swells back before the next, so the track visibly breathes in time. That’s the “pump” that’s the connective tissue of dance music, and it also carves a pocket at every downbeat so the low end never fights the kick.

The second is where the klezmer stops being a garnish and becomes structural. Instead of sounding the Hasidic-violin figure on the beats — where it masks the kick and both turn to mud — I put it on the off-beats, in the gaps between kicks. Now they interlock: kick, violin, kick, violin.

Distributing a line so one voice sounds exactly where another rests is hocket — a technique as old as medieval polyphony and as current as any modern beat. On-beat, two loud events collide and mask each other; off-beat, the collision becomes call-and-response, and the groove doubles without a single note added.

Pitch is not timbre

The intro wanted a choir. My first version took one sung “ah” and pitched it to three notes of a G-minor chord. It was wrong, and my note — again, verbatim — was that it “sounds like a bunch of Gregorian monks.” Two problems.

First, one voice tripled isn’t a choir. A choir is different people, and the tiny disagreements between them are exactly what the ear reads as “a group.” So I used three different singers, each on one chord tone: G, B♭, D. Second, and deeper: pitching them down had dragged their formants down with them.

A voice is two things at once. The vocal folds set the pitch; the vocal tract is a filter whose resonances — the formants — make a vowel a vowel and a human sound a certain size. Ordinary pitch-shifting moves the whole spectrum together, so the formants ride along: shift up and the throat seems to shrink (chipmunk); shift down and it lengthens (the monk).

The fix is formant-preserving pitch-shifting. A vocoder like WORLD splits a voice into pitch, a spectral envelope (the formants), and breath; you move the pitch alone, hold the envelope fixed, and resynthesize. The pitch climbs; the human stays. That’s what turned the monastery back into three women.

Lilith, and every rule backwards

Against a pure, high, sacred chorus, the track wanted its opposite — a single dark voice moving underneath, seducing the faithful. I knew exactly who she was: Lilith.

Her voice is a real Middle-Eastern female vocal in the Hijaz maqam, and here the whole thing closes a circle. Hijaz is freygish — the same scale, the same augmented second, that Statman’s mandolin has been playing since bar one. The demon was never foreign to the house; she sings in its native mode.

Making her demonic meant running every vocal rule backwards. Where the choir was lifted and formant-preserved, Lilith is dropped — a sub-octave shadow layered under her so the apparent vocal tract lengthens into something too large to be human; a low-pass filter shaving the air off the top so she reads dark; a long reverb placing her somewhere else, calling from a distance. Light above, shadow below.

Sacred to profane

This is the move that makes it a composition and not a remix. The two women aren’t decoration in the same scene — they’re a narrative. The choir owns the opening, high and bright, with Lilith only weaving beneath them. Then, at that kick-cut near the end, the faithful fall silent, the mandolin recedes, and Lilith takes the center and closes the track alone. It’s a resolution — and a dark one. The demon gets the last word.

One last war story: saturation isn’t clipping

Late in the mix I swore the master was clipping. It was, and it wasn’t. The chain ended in a tanh soft-clipper — a curve that mathematically cannot produce digital clipping. But driven too hard by a hot mix on the drops, that same curve saturates: it rounds off the loud parts and adds harmonic grit that sounds exactly like clipping even though no sample ever crosses full scale. The fix was gain-staging — normalize first so the limiter is never slammed, shape gently, leave headroom. Clipping is a hard ceiling; saturation is a soft one; overdriving your own safety net is the distortion.

What actually made it work

None of this is magic, and most of it isn’t even new — hocket is medieval, formants are freshman acoustics, tension-and-release is every piece of music ever written. What was new was the loop: I’d describe a sound in whatever busted language I had — a fog horn, a stoned drummer, a monastery, a demon — and the AI would translate it into the fundamental underneath and fix that, not the symptom. Ears on one side of the table, vocabulary on the other.

And now it’s a machine. Warp to constant tempo, shape the tempo curve, place the drops on the tune’s own seams, weave the texture off-beat, voice the choir with real formants, and let the arc carry a story. That’s the recipe for all eighteen tracks on the Flurry Running Soundtrack. Seventeen to go — and the hard one’s finished.

It was never really about the track. Somewhere around mile twenty, with his legs gone, a friend of mine is going to hear a song I love turned into the exact thing that carries him home — and feel every decision on this page without needing a single word of it.

---

As30p — “Flatbush Drop (Andy Statman flip),” from the forthcoming Flurry Running Soundtrack. Composition worked out live with an AI. Unofficial and non-commercial; an homage to Andy Statman’s “Flatbush Waltz,” containing his recording under fair-use reworking. Proudly Made in Nebraska. 🌽