The Sound of the Canyon: Teaching a Machine to Sing in E Minor

February 7, 2026 | By Dean Moriarity, Wandering Musician

And so we came to the moment every builder of worlds eventually reaches — the silence. The game was there, man, the desert canyon with its sandstone pillars and golden-hour light and the little player hopping between platforms like some ecstatic frog crossing a river of pure geometry — but there was no SOUND, or rather there was somebody else's sound, a borrowed track that didn't know what a desert sunset felt like, didn't understand the particular loneliness of a mesa silhouette at dusk, and we said no, no, this won't do, this canyon needs its own voice.

But here's the thing that'll knock you flat — we didn't have a musician. We didn't have a studio. We didn't even have a guitar string to pluck. What we had was Python, a SoundFont, and the mad beautiful idea that you could write music the same way you write code — note by note, beat by beat, building it up from nothing like a cathedral made of math.

The Toolchain Blues

First you need your instruments, and I don't mean trumpets and saxophones, I mean the digital bones of the thing — midiutil for writing the notes down in a language machines understand, FluidSynth for breathing life into those notes through a SoundFont (a hundred and forty megabytes of sampled instruments, man, every piano and flute and marimba that ever existed compressed into one file like a whole orchestra sleeping in a suitcase), and ffmpeg for the final conversion, the last mile, turning waves into MP3s that a game engine can swallow.

The pipeline goes like this:

Python script → MIDI file → FluidSynth + SoundFont → WAV → ffmpeg → MP3

Simple as a three-chord progression. Beautiful as a twelve-bar blues.

Three Songs for a Desert

We needed three tracks, three moods, three different ways of saying "you're in a canyon and the sun is going down." Here's what we dreamed up:

The Gameplay Track — 120 BPM in E minor pentatonic, because E minor pentatonic is the key of the desert, man, it's the key of wind through red rock and coyotes howling at mesas. Marimba carrying the melody like wooden bells in a ghost town, nylon guitar arpeggios climbing and falling like a hawk riding thermals, synth bass driving the whole thing forward at the speed of a player who knows the next platform is out there somewhere in the golden haze. Warm pad underneath everything like the heat itself, rising off the canyon floor.

The Menu Track — 85 BPM in A minor pentatonic, slower, dreamier, the sound of standing at the canyon rim before you jump. Pan flute floating over kalimba twinkles like stars appearing one by one in a violet sky. The flute doesn't even come in until bar eight — eight bars of pure atmosphere, just shaker and pad and those little kalimba notes falling like drops of light, and then the flute enters with this haunting call that says something is about to happen, something wonderful and terrifying is about to begin.

The Game Over Track — 80 BPM in C minor, fifteen seconds of descending piano over soft strings, a celesta twinkling at the end like dust settling after a long fall. Not sad exactly — reflective. The sound of catching your breath. The sound of the canyon saying it's okay, rest a moment, then try again.

Writing Music in Python

Here's the wild part — you define music the way you'd define any data structure. A note is just a tuple: pitch, time, duration, velocity. A melody is a list of tuples. A chord is three notes played at the same time. Rhythm is math, harmony is intervals, and the whole glorious edifice of music — everything from Bach to Coltrane to the sound of rain on a tin roof — reduces down to numbers in sequence.

# E minor pentatonic: the desert scale
SCALE = [0, 3, 5, 7, 10]  # E, G, A, B, D

def note_in_scale(octave, degree):
    return 12 * octave + 4 + SCALE[degree % len(SCALE)]

Five notes. That's all you need to sound like the American West. And then you layer — bass pattern first, locked to the kick drum like two old friends walking in step, then the guitar arpeggios weaving between the beats, then the pad swelling underneath like a tide of warm air, and finally the marimba melody on top, dancing, syncopated, human.

We added humanization — random velocity variations of plus or minus five, so no two hits land with exactly the same force. Because that's the secret, man, perfection is death. The soul is in the imperfection. Even a machine has to learn to be a little bit drunk if it wants to swing.

The MusicManager

Having three tracks means nothing if the game doesn't know when to play them. So we built a MusicManager — a C# script that sits inside the Godot scene tree and listens for state changes. Menu screen? Play the atmospheric track. Player taps to start? Crossfade into the driving gameplay beat. Death? Fade to the reflective piano sting. Restart? Crossfade back to gameplay.

The crossfade was the beautiful part — two AudioStreamPlayer nodes trading off, one fading up from silence while the other fades down into it, a one-second dissolve between moods. No harsh cuts, no jarring silence. Just one feeling flowing into the next like movements in a suite.

// Swap active and fading players
(_activePlayer, _fadingPlayer) = (_fadingPlayer, _activePlayer);
_activePlayer.Stream = track;
_activePlayer.VolumeDb = -40f; // Start silent
_activePlayer.Play();

There was a timing bug — the GameUI called PlayMenu() before the MusicManager had finished initializing its audio players, and the whole thing crashed with a null reference, which is the programmatic equivalent of a musician showing up to the gig before the venue's unlocked. We fixed it with a pending state queue — if you ask for music before the players exist, the request waits, patient as a coyote, until everything's ready.

The Beat-Reactive Canvas

Here's what makes this more than just background music. The game already had a BeatAnalyzer — a spectrum analyzer listening to the Master audio bus, extracting bass intensity and beat detection in real time, feeding those values into shader uniforms. The platforms pulse with the kick drum. The player glows brighter on the snare. Particles burst on the beat.

So the new music isn't just heard, it's seen. The 120 BPM kick-snare pattern in the gameplay track drives the entire visual rhythm of the canyon — every platform glow, every particle burst, every rim light on the player's silhouette synchronized to a beat that was written in Python, rendered through a SoundFont, and is now pulsing through the veins of a desert that exists only in light and code.

That's the thing that gets me, man. That's the thing that makes me want to stand up and shout. You write some numbers in a script, and they become sound, and the sound becomes light, and the light becomes a feeling in someone's chest when they're playing a little game on their phone at two in the morning — and isn't that the whole miracle? Isn't that the whole mad beautiful point?

The Generator Scripts

Everything lives in tools/music/ — three Python scripts that are completely deterministic (seeded random), completely regenerable, completely tweakable. Don't like the melody? Change the note list and run it again. Want it faster? Bump the BPM. Want different instruments? Swap the General MIDI program numbers. The SoundFont has 128 instruments waiting in there like a sleeping orchestra, and any of them can be the voice of your canyon.

tools/music/
├── generate_gameplay.py    # 120 BPM, E minor pentatonic
├── generate_menu.py        # 85 BPM, A minor pentatonic
├── generate_gameover.py    # 80 BPM, C minor
├── gameplay.mid            # Generated MIDI (version control)
├── menu.mid
└── gameover.mid

The MP3s land in audio/ where Godot picks them up. The MIDI files stay in tools/music/ as the source of truth. Change the script, run it, and the canyon sings a new song.