Most synthesizers are built for people who already know synthesis. The interface is a wall of knobs labeled with words like "resonance", "attack", and "LFO rate" — terms that mean nothing to someone who just wants to make a sound they hear in their head. The learning curve isn't steep; it's vertical.
This project started from a different question: what if you could describe a sound the way you'd describe it to a friend? Say "warm bass pad" or "icy bell that shimmers" or "punchy 808 with a long tail" — and the synth figures out the rest. That's the core idea. The AI doesn't replace the knobs; it gives you a starting point you can hear immediately, then you tweak from there.
Under the hood, the synth uses GPT-4o-mini with a heavily engineered system prompt — a 3,000-word sound design reference guide that teaches the model the physics of waveforms, the relationships between filter types and harmonic content, and 20+ real preset recipes it can use as building blocks. When you type "dreamy keys," the AI isn't guessing — it's blending the envelope of a vibraphone with the detuned oscillators of a lush pad and the reverb profile of a concert hall.
The entire synth runs client-side in the browser with zero server infrastructure. The Web Audio API handles real-time DSP — oscillators, filters, envelopes, effects chains, a brick-wall limiter — all at 44.1kHz with sub-millisecond latency. MIDI keyboards plug in via the Web MIDI API with hot-plug detection and sustain pedal support. You can export anything you make as a WAV file or save it as a .syn patch file to share with others.
The goal is a synth that a professional producer would use alongside their DAW, but that a fourteen-year-old with no music theory could pick up and start creating with in under thirty seconds.