Procedural Lofi — build a soundtrack in code, not in a sample pack

This skill is the end-to-end recipe for a lofi bed (and its sibling ambient sounds: rain, keystrokes, door chimes, printer, water cooler…) that is synthesized at runtime from constants — zero audio files, zero royalties, and no two minutes ever sound the same. It was proven out on a real shipping product (an animated pixel-art office that plays ambient sound scaled by how busy the on-screen agents are), then distilled here.

Two documents ship alongside this one:

• reference/LOFI-BIBLE.md — the parameter tables. Harmony (chord grammar, voice-leading, register clamps), groove (swing %, drag ms, velocity curves), per-voice sound design (pad / bass / EP keys / sparkle / drums / texture / tape chain), mix targets (band shares, HPF strategy, loudness), and generative lessons from prior art. Every number is cited and, where it conflicts with a measurement, the measurement wins.
• scripts/ — the Python (numpy) pipeline: analyze_*.py (fingerprint a reference), synth_audition.py (a runnable numpy synth that produces .wav auditions), and export_score.py (freeze a good take into a constants table for the port).

Read those two when you reach the step that needs them. This file is the map.

The core idea (why code, not samples)

A sampled lofi loop is: (a) a licensing liability, (b) heavy to bundle, and (c) audibly repetitive — the human brain catches a loop seam on the 3rd or 4th pass. Synthesizing it solves all three:

• Legal: acoustic parameters are not copyrightable. You measure a reference to get target numbers (this band has 55% of the power, the centroid sits at 150 Hz, drops land 13 dB above the bed). You never keep or ship the reference bytes. What ships is your own oscillator code hitting those numbers.
• Size: a few kilobytes of constant tables + a synth function vs. megabytes of PCM.
• Never repeats (in practice), from three stacked tricks: (1) the frozen musical composition — pad, drums, and the melody (keys + sparkle) — loops in lockstep, but the loop is made long enough that its repetition doesn't fatigue (this project doubled its day loop from 4 to 8 bars precisely because a short looped melody was audibly repetitive — "loop the bed" does not mean "make it short"); (2) the genuinely-stochastic layers — rain drops and keystrokes — are scattered fresh at runtime from a seeded RNG, laid over the loop; (3) a busy-ness stem mixer fades whole stems (drums, keys…) in and out by how active the scene is, so the arrangement keeps changing even though the notes don't. The result reads as "never the same two minutes" without regenerating harmony. (Regenerating melody live is possible but it's the hard, optional part — the shipped product froze the melody and leaned on 1–3 instead.)

The one law of the whole method:

Measurement is the machine's ears; taste is the human's. You (or your tooling) cannot reliably hear whether it sounds good — you can only measure proxies (band energies, onset rate, dB deltas) and drive them to a target. A human listens once per major revision and gives a yes/no. Pick your measurable proxy carefully, iterate on it alone, and hand over a finished audition — don't ask the human to babysit each tweak.

The pipeline (five phases)

Phase 0 — Audition in a scripting language first (numpy), NOT in your ship language

Prototype the whole sound in Python/numpy where the write→hear loop is seconds. Only port to Rust/C++/wasm once a human has ratified the sound. scripts/synth_audition.py is a worked example: deterministic numpy that writes audio-demos/*.wav. Iterating synth recipes in a compiled language first is the classic time sink.

Phase 1 — Reference-fingerprint each sound

For every distinct sound (the lofi bed, rain, typing, each one-shot):

1. Get a beloved reference. An owner-supplied reference beats a "community ideal" every time — build to what they love, not to what a forum says lofi should be. (yt-dlp / curl for analysis only; for an endless live stream, yt-dlp -g gets the HLS URL, then ffmpeg -t 180 grabs a finite slice.)
2. Confirm the CHARACTER before deep-matching. One sentence — "gentle rain or heavy downpour?" — saves three wasted versions. On this project, three rain versions were built to the wrong reference (a heavy wash) before the owner clarified they wanted gentle rain.
3. Fingerprint it. Measure 9 octave-band energies + spectral centroid + rolloff (analyze_rain.py / analyze_typing.py). For anything rhythmic or event-bearing ALSO measure the temporal fingerprint — onset rate, inter-onset-interval spread, per-stroke decay, and the dB level of foreground events vs. the bed (analyze_drops.py). Spectral averages are blind to events: rain's audible drops don't show up in an averaged spectrum at all, only in the temporal pass.

Phase 2 — Shape synthesis to the measured curve

Build your oscillators/noise-shapers and drive their parameters until a re-measurement of your output lands within a few percentage points of the reference fingerprint. The LOFI-BIBLE.md gives you the starting parameter values per voice; the fingerprint tells you which way to push them. Search the literature for the physics (Minnaert resonance for a water glug, inharmonic bar modes 1 : 2.76 : 5.40 for a chime, tape wow/flutter + head-bump for the lofi chain) but tune to the reference, not to the physics ideal.

Phase 3 — Human LISTEN gate → freeze the realization

When the numbers converge, a human auditions once (afplay file.wav on macOS, or hand them the wav). On yes:

The RNG was the composer. Freeze the one take they blessed.

The generative script drew notes/velocities/timings from a seeded RNG. That one seed's output is what got ratified — so capture its exact event stream into constant tables (export_score.py → a .rs / .h table), plus a full-table checksum. Do NOT re-run the RNG in production and hope; a later library bump silently redraws and you ship a different, un-ratified take. The freeze is the contract. (Subtlety: your exporter must reproduce the exact draw order of the original — argument-evaluation order, nested draws inside a choice() — or the frozen table desyncs from what was auditioned.)

Phase 4 — Port + ship (synthesize at launch)

Port the numpy synth to your runtime language reading the frozen tables. Key engineering:

• The port must be byte-identical to the audition. Keep the ratified fingerprint/ checksum tests; passing them verbatim after the port is your oracle that nothing shifted.
• No wall-clock reads inside the audio math. Pass dt / now in as parameters. On wasm especially, SystemTime::now() isn't available — and a backgrounded tab that jumps the clock will otherwise ramp-snap your crossfades and burst-replay every queued event (the "stall-clock" bug). Clamp big dt gaps.
• Loop the frozen composition, scatter only the stochastic layers. All the musical stems (pad/drums/keys/sparkle) tile in lockstep on one loop length — make that loop long enough to not fatigue; then scatter the truly-random layers (rain drops, keystrokes) fresh each frame and fade whole stems by busy-ness (see "Never repeats" above). Tiny RAM: one copy of each stem loop.
• Keep the sub band clear. The sub-bass register must belong to the bass alone or the low end turns to mud. Two ways to get there: the textbook one is to high-pass every non-bass stem ~140 Hz (LOFI-BIBLE §4); the cheaper one this project used is to voice the other stems out of that register in the first place (mid-register EP plucks, and a day track with no sub content at all) so there's nothing to filter. Either way, keep ONE texture stem carrying all the "medium" noise — per-stem hiss stacks (four stems each with a little = +6 dB of mud); texture sits 25–35 dB below the music.
• Perceptual volume. Map a user volume slider as amplitude = user² (loudness is logarithmic). Linear volume feels "still too loud at 5%" — the classic trap.

The failure catalog (bugs this method hit, so you don't)

• Reference outranks literature. A "thock" keyboard sound built to the ASMR-community ideal measured the opposite of the owner's clacky reference. Search for physics; tune to the reference.
• Phantom-octave measurement bug. Fingerprints extracted from the listening artifacts (stereo wav, soft-clipped) read every frequency halved (interleaved stereo read as mono = 2× time-stretch). Pin spectral references on the raw float synthesis chain, never on the exported/played wav.
• Spectral averages can't see events. Rain drops, a snare ghost, a keystroke accent — measure them temporally (level vs. bed, onset rate) or you'll "match" a reference and still be missing its soul.
• Head-only crossfade ≠ a seamless loop. Fading the start of a loop back over its end leaves the wrap audible. A real seamless loop synthesizes a genuine continuation past the loop point and blends that in.
• Two frozen copies of one truth desync silently. If a kick-times table duplicates the drum table's kicks, add an equality test or a later edit to one drifts from the other with no error.
• Control plane off the data plane. A mute/volume press must not ride the same queue as audio frames — a synthesis burst can saturate the queue and eat the keypress, so the beds fade in unmuted while the UI says muted. Put mute/volume on an atomic flag the audio thread reads, separate from the frame channel.
• Empty-room-uncanny. A truly silent "idle" state feels broken, but a full bed under nothing feels wrong too. Keep the register (don't octave-drop), voice harmonically, sparse EP notes, and let the texture bed carry the quiet — don't drop to pure silence.

Quickstart for a fresh project

1. Copy reference/LOFI-BIBLE.md and scripts/ into your repo (or just read them).
2. Install: python3 -m venv .venv && .venv/bin/pip install numpy scipy (+ yt-dlp, ffmpeg for grabbing references).
3. Pick and character-confirm a reference. Fingerprint it with the matching analyze_*.py.
4. Adapt synth_audition.py's voices toward the fingerprint; re-measure to convergence.
5. Human listens once. On yes, export_score.py freezes the take + checksum.
6. Port to your runtime; keep the fingerprint tests as the byte-identity oracle; synthesize at launch, loop the frozen composition, scatter the stochastic layers, gate stems by busy-ness.

(The scripts/ are the real working prototypes from this project, kept as a concrete starting point — not a turnkey CLI; synth_audition.py runs standalone, and export_score.py imports phase2_audition.py — both are bundled. Adapt the voices to your own reference.)

The whole discipline in one line: fingerprint a reference you love, drive your own synthesis to the numbers, freeze the one take a human blesses, then let a long-enough loop + runtime-stochastic layers + busy-ness stem-gating keep it from ever sounding the same.

The GENERATOR loop (pixtuoid's Rust composer — iterate in minutes, not arcs)

Once a project graduates from frozen takes to a theory-constrained generator (pixtuoid-scene/src/audio/compose.rs + synth::gen_beds), the iteration loop changes shape: the LISTEN gate becomes statistical (blind-audition a batch of seeds; all acceptable = the GENERATOR is ratified — a dud = tighten a constraint and re-batch), and every taste axis is a data table or named const with ONE home:

Add-an-instrument checklist (executed once for the Pluck lead — repeat verbatim):

1. Write the voice as a pure note fn in synth.rs (fn my_voice(midi, dur_s, vel) -> Vec<f32>), physics-first then tuned by ear/fingerprint (Phase 1-2 of this skill still apply).
2. Add a compose::LeadVoice variant + its arm in synth::lead_voice_fn (the ONE dispatch).
3. Give it a draw weight at the END of compose() — the voice draw is deliberately LAST in the seed stream so new voices never redraw an already-blessed seed's notes.
4. Extend the distribution property test; run the compose suite (fast — no synthesis).
5. Render a batch (cargo run --release -p pixtuoid-scene --example lofi_audition -- --seeds N; --solo sparkle isolates the lane) → owner listens → tighten or ship.

Mix LANES stay instrument-blind (StemLevels/mixer/players never learn about voices) — a lane is a busy-ness ROLE, an instrument is a timbre WITHIN it. That split is what keeps "blend in another instrument" a minutes-scale edit instead of an arc.