two engines, one score

I’ve been building aether-synth — a code-built synth studio. The idea is simple: you write music as a small DSL, and you hear it immediately.

tempo 112

track lead {
  osc sawtooth
  adsr 0.01 0.18 0.55 0.25
  filter 2200 6
  gain 0.55
  seq 16 [C4, _, E4, G4, _, A4, _, C5, _, A4, _, G4, _, E4, _, D4]
}

The browser parses that and plays it through Web Audio in real time — oscillators, ADSR, a lowpass filter, delay, reverb, a 16-step sequencer, all synthesized live. No samples. Change a number, hear the difference instantly. That’s the sketchpad.

The printing press

But real-time Web Audio is a sketch. For an actual record, I wanted an engine that is exact — deterministic, anti-aliased, the kind of thing that could one day live inside a VST. So there’s a second renderer: a zero-dependency C++17 engine that compiles the .msong score format to a stereo 16-bit/44.1kHz WAV, about 52× faster than real time. PolyBLEP anti-aliasing, RBJ biquad filters, a Schroeder reverb, 32-voice polyphony per track. The audio path is real-time-safe: fixed 512-frame blocks, zero heap allocation after prepareToPlay. It’s the printing press to the browser’s sketchpad.

The Node API in the middle is almost embarrassingly thin — it writes your score to a temp file, spawns engine/msynth render, and streams the WAV back. The interesting work is the bridge: converting the browser DSL to .msong.

The test that paid for itself

The round-trip test — parse DSL → convert → render with the real C++ binary — caught a genuine bug on the first run. The .msong README documented parameters with quotes (adsr "0.01 0.18 0.55 0.25"), but the actual parser splits on spaces and never strips quotes. One line in a test, and the documentation lied to me. The test now renders through the real engine, so the format can never drift again without a failure.

That’s the whole lesson, really: don’t test your converter against a mock. Test it against the thing it’s supposed to talk to.

Why two engines?

Because they’re for different moments. The browser engine is immediate — you hear a change the moment you type it. The C++ engine is permanent — deterministic, exact, and fast enough to render a full piece while you make coffee. Same score, two characters. The sketchpad and the printing press. I want both.

The repo is SaddleDove/aether-synth — MIT, both engines, all tests green.