@audio/midi-render

MIDI → PCM — General MIDI synth render, no soundfont, no samples

import render, { GM, voiceFor } from '@audio/midi-render'

A synth-GM: every one of the 128 General MIDI Level 1 programs and every percussion key (35-81) maps to an @audio/synth-* generator — FM (Chowning 1973), Karplus-Strong plucked strings (Karplus & Strong 1983), and modal resonator banks (Fletcher & Rossing, The Physics of Musical Instruments, 1998) — instead of a sampled/SF2 bank. Parses with @audio/midi-parse, allocates polyphony per channel with @audio/synth-poly (note-stealing), mixes to stereo. Zero asset weight: the whole instrument set ships as code.

import render from '@audio/midi-render'
import { readFileSync } from 'fs'

let smf = readFileSync('song.mid')
let { channelData, sampleRate, duration } = render(smf)
// channelData: [Float32Array, Float32Array] (L, R) — write to WAV/whatever with @audio/encode-wav

// override an instrument, or inspect the map
import render, { GM, voiceFor } from '@audio/midi-render'
import bell from '@audio/synth-fm'

render(smf, { programs: { 8: (freq, opts) => bell(freq, { ...opts, index: 12 }) } })
console.log(GM.programs[0].name)         // 'Acoustic Grand Piano'
console.log(voiceFor(0)(440, { fs: 44100, duration: 1, velocity: 0.8 }).length)

The GM map

Every program/key is covered — render() never falls silent on a valid GM number, and an out-of-map percussion key gets a short click rather than nothing.

Velocity maps to amplitude through a v^1.5 power curve (perceptually closer to a loudness control than raw linear velocity), applied once before synth-poly's own linear gain stage.

MIDI control handling

• Program changes — honored per note, per channel (mid-track changes apply to every note from that point on).
• CC7 (channel volume) — read per note (the value in effect at the note's start time); default 100/127 per the GM System Level 1 spec's default controller value.
• CC10 (pan) — equal-power, read once per channel at the value in effect at the channel's first note. Mid-channel pan automation is not applied per note — each channel renders to one mono buffer before panning, so a single pan value applies to the whole channel. Default center (64).
• CC64 (sustain pedal) — extends a note's held duration until the pedal releases (or the end of the piece, if never released). Applied per note, before synthesis.
• Pitch bend — read once per note, at the value in effect when the note starts, and applied as a constant per-note transposition (bendRange semitones at full deflection). Bend automation during a held note is not applied — there's no per-sample pitch modulation. Custom RPN 0,0 bend-range messages are not parsed; only the bendRange option changes the default ±2 semitones.
• Percussion (channel 10) — program changes are ignored; the fixed GM1 percussion map always applies (matching GM Level 1, which defines a single kit).

Performance

Rendering is deterministic per (instrument, freq, duration) — repeated notes (very common in real MIDI — the same pitch/duration/instrument recurs constantly) reuse a cached Float32Array within one render() call instead of re-synthesizing. A 2-minute, 8-channel, ~1500-note file renders in ≈4s on a laptop CPU (measured in test.js). Each channel is rendered and mixed down before moving to the next — the renderer never holds every note's audio in memory at once.

Use when: you need audio from a Standard MIDI File and don't want a soundfont/sample-bank dependency — previews, MIDI-to-audio-in-the-browser, test fixtures, procedurally generated music. Not a substitute for a real soundfont/sample player when sample-accurate GM-compatible timbre matters (a game engine's music, a DAW export) — the instrument bank here is a from-scratch synthesis approximation of each GM voice, not a recording of one.

Part of @audio/midi — the midi family umbrella.

MIT © audiojs