The whole synthesizer is one file of C. It compiles to 165 KB of wasm, lives on the audio thread, and does not glitch. Ever.

Browser audio engines synthesize on the main thread, and the main thread has other plans — one garbage-collection pause, one fat framework commit, one backgrounded tab, and the music cracks. Every fix is an apology: bigger buffers, simplified mixes for phones, reverb rationing on mobile.

zaltz deletes the problem instead of managing it. Oscillators, filters, envelopes, the sampler, FDN reverb, delay lines, sidechain compression — all of it is one file of freestanding C, no libc, no SharedArrayBuffer, no cross-origin headers, rendering inside an AudioWorklet where your UI thread physically cannot reach it. The page can stutter, sleep, or die mid-animation. The music does not.

Hear it: zaltz.klappn.com — type a pattern, it renders live.

zaltz is the engine under every song at Klappn. It replaced superdough — Strudel's sampler/synth layer — after being measured against it, case by case, until a golden-gate harness couldn't tell them apart: envelope correlation, level, brightness, per feature, against superdough's own offline render. Then an 8-song, 70-loop corpus of real music was A/B'd by ear before it became the default.

The day it shipped, Klappn deleted its entire mobile apology layer: the AI-generated "lite" rewrites of every song for phones, the reverb caps, the per-device mixes. Phones now play the same full mix as desktops, because the render doesn't live where phones are slow.

• engine/zaltz.c — ~1,600 lines. Band-limited wavetable oscillators (triangle at sample-exact 90° phase — a phase-blind test suite hid that bug once, never again), four noise colors, ADSR + filter envelopes, ladder/12dB/24dB filters, a growable sample store, looping soundfont voices, and per-orbit buses: FDN reverb, delay, sidechain duck, phaser, waveshaping.
• dist/zaltz.worklet.js — the AudioWorklet host. Events are written straight into engine memory (zero allocation per event on the audio thread — worklet GC pauses are audible, so there are none), sample uploads drain under a strict per-quantum budget, and an audio-thread clock posts outward so scheduling survives background-tab timer throttling.
• index.js — a thin, typed host API: boot, schedule, hush, crossfade-retire, orbit gains, sample upload.

import Zaltz from "zaltz";

const ctx = new AudioContext();
const z = await Zaltz.create(ctx);

// A bar of dub: bass on the beat, stabs on the off-beat ringing
// through a synced delay into the FDN reverb on orbit 2.
const t0 = ctx.currentTime + 0.1;
const spb = 60 / 122; // seconds per beat
const bass = [33, 33, 36, 31]; // A1 A1 C2 G1
z.scheduleAll(bass.flatMap((note, beat) => [
  { params: { s: "sawtooth", note, duration: spb * 0.9, attack: 0.005,
      release: 0.1, lpf: 500, gain: 0.8 },
    when: t0 + beat * spb },
  { params: { s: "sawtooth", note: note + 24, duration: 0.15, release: 0.12,
      lpf: 1500, room: 0.45, roomsize: 6, delay: 0.4, delaytime: spb * 0.75,
      delayfeedback: 0.5, orbit: 2, gain: 0.5 },
    when: t0 + (beat + 0.5) * spb },
]));

Params speak superdough's language — s, note, ADSR, lpf, room, delay, shape, orbit, duck — so anything that already talks to Strudel can learn to talk to zaltz in an afternoon. The pattern layer stays upstream; this package is the sound.

Every rule in the engine was paid for with a click somebody heard:

• Buses ramp, never step — including feedback-network coefficients. A retuned reverb glides its damping and size while the tail rings. Stepping a live coefficient is a click by construction, so it is impossible by construction.
• Nothing unbounded on the audio thread. Event writes are length-guarded, uploads budgeted per render quantum, memory growth chunked and rare.
• Parity is the floor, not the ceiling. Where the reference architecture glitches — a shared convolver regenerated mid-ring, per-event GC churn — zaltz deliberately diverges.
• The ear is the acceptance test. Metrics gate regressions. A human listening decides quality.

No SIMD (Safari 15 runs it natively), no SharedArrayBuffer (no COOP/COEP headers to negotiate), no main-thread synthesis (phones keep up). Memory starts at a few MB and grows only with what a song actually loads. Rebuild from source with npm run build — one zig cc invocation, no toolchain ceremony.

AGPL-3.0-or-later — zaltz is a derivative of superdough (the Strudel project), and it honors its ancestor's terms; see NOTICE.md.