@driftbox/engine

A TR-808, a TR-909 and two TB-303s, synthesised from scratch with Web Audio. No recorded samples — the analogue voices follow their circuit topologies, while the 909's digital cymbals use generated 6-bit/30kHz PCM and modal synthesis. The 303 includes a real 4-pole ladder filter running in an AudioWorklet.

Imports no React, touches no DOM, and does not know a sequencer UI exists. It is meant to be embedded — in a game, an installation, or your own sequencer.

npm install @driftbox/engine

import { DriftboxEngine, defaultSong } from '@driftbox/engine'

const engine = new DriftboxEngine(defaultSong())
await engine.start()        // from a click; browsers start audio contexts suspended
engine.audition('909.cp')   // a one-shot, for a game event

Browser only — it is Web Audio from top to bottom. It imports cleanly in Node (useful for tests and tooling) but needs a real AudioContext to make a sound.

A song is a value

import { defaultSong, encodeSong, decodeSong, type Song } from '@driftbox/engine'

A Song is plain JSON: patterns, an arrangement, kit settings and effect sends. So a soundtrack ships as an asset rather than as code, and decodeSong treats anything it is handed as untrusted — clamping what is out of range, filling what is missing, and returning null only for input that is not a song at all.

The four shipped patterns (haze, drift, neon, pulse) come with the package, so there is something to play before you have written anything.

Adaptive soundtracks

The chain is data, so a host can rewrite it as the thing it is scoring changes:

engine.song = { ...engine.song, chain: [{ pattern: 'haze', repeat: 4 }] }

Sections take effect at the next bar boundary, because the transport reads the arrangement one bar ahead rather than caching it.

Voices are pure functions

A voice is a pure function from its knob positions to a VoiceSpec — a description of oscillators, noise sources, envelopes and filters. A separate renderer turns that spec into Web Audio nodes, so the synthesis is testable without an audio device, a patch can be serialised, and the same spec renders into a live context or an OfflineAudioContext:

import { renderVoiceOffline, voiceById } from '@driftbox/engine'

const samples = await renderVoiceOffline(voiceById('808.bd'))

The 303s

A BiquadFilterNode cannot be a 303: two poles, linear resonance, no self-oscillation. This uses Huovilainen's nonlinear ladder model in an AudioWorklet — four poles, feedback past the point of oscillation, and a tanh inside the loop, which is what makes the resonance thick rather than a whistle on top. Measured: self-oscillates at full resonance, tracks the requested cutoff within 3% from 300Hz to 2.4kHz, and stays bounded however hard it is driven.

Slide and accent are the rest of it. A slide glides between two notes sharing one envelope; a paused step retains its pitch and can bend the following attack from silence, as ReBirth did. An accent drives level, filter envelope depth and resonance together, as the accent voltage does on the hardware.

If AudioWorklet is unavailable — an old browser, or a Content-Security-Policy that blocks blob scripts — the 303s fall back to a single biquad and sweep rather than squelch. engine.usingLadder says which you got.

Licence

MIT. The ladder filter follows the model published by Antti Huovilainen ("Non-linear digital implementation of the Moog ladder filter", DAFx-04); the implementation is original.

Not affiliated with, endorsed by, or connected to Roland Corporation. TR-808, TR-909 and TB-303 are Roland's trademarks and appear here only to describe which instrument each voice is modelled on. Every waveform is generated: there are no recorded samples, and no recordings or ROM data from any hardware were used.