What we build
Four product lines, one shared idea about where the hard work belongs.
Astral Software is a small independent studio. We make instruments, apps and installations, and the thing they have in common is that we build the layer underneath them ourselves rather than renting it.
That started with audio. Our first instrument ran in a browser, and the browser gave us roughly 25 ms of latency, 190 ms of pitch lag, worklet limits we could not design around, and a garbage collector that could interrupt a performance. So we wrote a native engine on JUCE and C++ — a host-agnostic core with a lock-free queue between the control thread and the audio thread — and every audio product we have made since sits on top of it.
The same principle repeats elsewhere. Stem separation is a model we are training rather than an API we call. AI-music detection is a classifier we built and measured, and we publish the blind-set number alongside the flattering one. The detection layer in our interactive installations is a deterministic OpenCV filtering stack, not a neural net — because it has to work in a venue, on modest hardware, with a projector shining into the same camera that does the tracking.
None of this is ideology about writing everything from scratch. It is that the hard part is usually the part everyone else abstracts away, and once you own it, the products above it get better for free.