Generative music for a 24/7 OpenPost Radio stream on YouTube and Twitch: work and focus music written as Strudel code, rendered offline into a track library, and streamed with a calm dithered scene.
compose (seeded Strudel code) → render (Node, no browser) → master (-14 LUFS FLAC)
→ library (local or S3/R2) → stream (director → crossfade → scene → ffmpeg → YouTube, Twitch)
devenv shell # or: nix shell nixpkgs#nodejs_24 nixpkgs#ffmpeg
npm install # also bundles Strudel for Node (dist/strudel.js)
npm run compose -- --style melodic --seed 42 # print the track's Strudel code
npm run compose -- --style melodic --seed 42 --render out.wav # render it
npm run library -- --count 40 # build 40 mastered tracks in parallel
npm run stream -- --output preview.mp4 --duration 120 # preview the stream as a file
npm run scene -- --at 2026-01-12T17:40 # still of the scene at a time (tmp/scene)
npm run stream -- --tour --output tour.mp4 # fast-forward through every place and kind of travel
npm run audition -- --styles melodic --seeds 1,2 # score excerpts with earshot (needs uv)
npm testEvery track is plain Strudel code. Paste library/tracks/<id>.strudel into
strudel.cc to hear it in the browser and edit it.
Four styles share one engine, kit and mix, modelled on the reference mixes analysed in docs/music/reference-analysis.md:
| style | label | bpm | character |
|---|---|---|---|
| melodic | melodic house | 118-127 | an eight-bar chord loop held over one sidechained sub, a short hook crossing the bar line |
| deep | deep house | 119-123 | one or two chords for a whole song, offbeat stabs, a plucked bass that avoids the kick, vocal chops |
| broken | broken beat | 128-134 | two kicks a bar off the grid, steady sixteenths, held ninth chords |
| slow | slow soul | 80-95 | electric piano leading a loose beat, a walking bass, pentatonic fills at phrase ends |
What keeps 8 hours from getting boring:
- Composition (
src/compose): every song is one harmonic loop in any key, voiced as rootless close chords with minimal movement. Forms are 8 and 16 bar blocks. Parts enter on block lines, a one-bar kick dropout closes a block, and a filter arc over the whole song does the work chord changes would otherwise do. Each seed makes one or two clear choices (held or stabbed chords, a hook or a falling run) rather than using every part. - Sound (
src/compose/instruments.js): presets for sub, bass, stabs, held chords and hooks, and a kit synthesised byscripts/build-kit.jsto the measured kick, clap and hat of the references. - Restraint: chords are four or five notes, one chord part and one melody at a time, rooms are short and dark, and upper parts leave gaps. Those gaps are what separate these songs from a wall of sound.
- Mix:
npm run calibratemeasures every preset on a reference phrase and stores gain factors incalibration.json; styles set levels per role. Any preset combination therefore lands at the same balance. Fixed effect buses per orbit keep reverbs constant. The kick pumps chords and bass, as much as 8 dB in the house styles. - Targets:
tools/reference-analysismeasures renders and references the same way. House mains sit near 90% of energy below 400 Hz with a 250-350 Hz centroid, a mid-band floor well below the peaks and a crest factor near 11.5 dB. Slow soul is darker, about -24 dB above 2 kHz. - Listening:
npm run auditionrenders the opening of a few seeds per style and scores them with earshot: how well each matches its style's description against "cheap synth demo", "corporate" and "video game" alternatives, and how produced it sounds. Results land intmp/audition/<label>, so changes can be compared before and after. Treat the scores as a guard against regressions, not as taste.
Intros and outros are always beatless, so the stream can crossfade between any two tempos.
src/engine runs Strudel's own evaluator and superdough on
node-web-audio-api. The
library needed several fixes, all in src/engine and
scripts/build-strudel.js:
- Tracks render in 4-bar chunks, each in a fresh
OfflineAudioContextwith 6 seconds of pre-roll. A single context is quadratic in length, and adding voices mid-render withsuspend()intermittently froze the output. Chunks are checked for frozen output, and each seam must match its neighbour sample-for-sample; a failing chunk is rendered again. - superdough's per-voice AudioWorklets (LFOs, shape, distort, crush, supersaw) are replaced with equivalent native nodes. Worklet nodes are never freed offline in node-web-audio-api.
- Reverb impulse responses, noise, crackle, FM modulator phase and supersaw phases are deterministic, so separately rendered chunks agree.
- Sidechain ducking is written as direct gain automation.
- WAV files the native decoder rejects fall back to ffmpeg.
Renders run 3-40x faster than real time per core. The library builder runs one
track per CPU core. Sample maps and samples are cached in cache/, so renders
after the first work offline.
Fidelity against strudel.cc (Chrome) was A/B tested. Oscillators, filters, samples and soundfonts match. FM timbres differ slightly because modulators now start in phase with each note.
npm run stream (src/stream/main.js):
- Director: melodic and deep house trade places, with broken beats and slow soul as a change of pace. Style runs are a few tracks long, keys stay Camelot-compatible, tempos stay close, nothing repeats until about 60% of the library has played, and late nights lean calmer.
- Mixer: equal-power crossfades across the outgoing outro and the incoming intro (6-24 s).
- Scene (
src/stream/scene): a 640x360 dithered landscape in parallax, built from long painted plates and recoloured each frame by one light model. A fox with a postbag travels right and is in motion at least 95% of the time: it walks, finds a bicycle, rows lakes, sails coasts and rides a broom, stopping only for a quick workout or a telescope look. Friends pass the other way without stopping. The wall clock drives a 40-minute world day and the weather (cloud, rain, fog, snow, wind); distance drives the country (meadow, valley, lake, alpine, forest, coast) and the season, so the fox walks into the snow. The only text is the two top chips. - Encoder: one ffmpeg process upscales with nearest neighbour (2x for
720p, 3x for 1080p pixel art), draws the
openpo.st/radiochip (top left) and now-playing text (Geist Mono), and encodes H.264 + AAC once for every output. Several outputs (YouTube and Twitch) share the encode through ffmpeg's tee muxer, each behind its own fifo, so a platform that drops reconnects without stalling the others. If ffmpeg itself exits, the stream restarts it.
library/ holds tracks/<id>.flac, its metadata .json and source
.strudel, plus catalog.json. Tracks are 16-bit FLAC at -14 LUFS, about
35 MB for five minutes.
RGO_MUSIC_REMOTE=r2:bucket/library scripts/upload-library.shThe stream reads either a local directory (RGO_MUSIC_LIBRARY) or a base URL
(RGO_MUSIC_LIBRARY_URL, for example a public R2 bucket). With a URL it
prefetches ten compatible upcoming tracks in the background. Its cache
protects current, fading, and queued audio while pruning older tracks.
The live sender runs on rgo-vps. The NAS is storage only. To minimize CPU
and RAM, prepare the complete video and audio on a workstation, then let
src/stream/relay.js forward H.264/AAC packets with FFmpeg -c copy.
There is no scene renderer, decoder, synthesizer, or video encoder on the VPS.
node src/cli/program.js --hours 8 --fps 30This publishes closed four-minute MPEG-TS chunks and an atomic manifest to
/volume1/media/openpost-radio-music/programmes/ on the NAS.
Two FIFO readers stream video directly into temporary NAS files. Each file
is renamed after its upload closes. The local work directory holds only
small metadata files and pipes, so preparation needs little local disk space.
Set RGO_MUSIC_REVISION to record the source revision in the manifest. The
initial programme is 1280x720 at 30fps, 3 Mbps H.264 plus 160 kbps AAC, with
two-second keyframes. Its music crossfades and visualizer are baked into the
video. The eight-hour programme repeats, with a 12-second audio fade and a
short dither at the loop boundary. Fresh compositions and seasonal changes
require preparing and publishing another programme, then restarting the relay.
The live sender does not generate new music or weather indefinitely.
For a genuine 1080p60 programme on macOS, using hardware encoding:
node src/cli/program.js --hours 8 --fps 60 --scale 3 --bitrate 6000k --encoder h264_videotoolboxPreparation draws every frame at 60fps. It does not duplicate a 30fps movie.
The programme manifest becomes current only after all chunks are published.
Warm the relay cache, then restart openpost-radio.service to adopt it. Run
scripts/prewarm-relay-cache.sh first: the relay only starts FFmpeg once its
11-chunk window is on disk, so a cold-cache restart leaves the YouTube ingest
with no data for minutes and YouTube responds by looping its buffered window on
air. With the cache warm the restart reconnects in seconds. At the 6 Mbps video budget,
allow about 2.1 GB of VPS cache for the current chunk plus ten ahead.
Create or replace the YouTube broadcast first, then restart the relay. A broadcast that attaches to a sender session already hours old leaves YouTube's ingest unable to advance: it keeps the connection and replays its ~8-second prebuffer forever, so audio and video play back stacked on themselves. A fresh session starts at zero-based timestamps together with the broadcast.
The VPS keeps the current chunk and ten ahead, at most eleven chunks. That gives about 40 minutes of NAS-outage headroom after the buffer fills, using roughly 1.1 GB of bounded cache instead of the entire library. A dedicated restricted SSH key can read only radio assets, with a pinned NAS host key. Only two downloads run at once. Startup waits until the lookahead has been published. A reconnect seeks to the current wall-clock position in the programme.
The service and encrypted credentials are owned by
~/.config/home/modules/services/openpost-radio/default.nix. The module packages a reviewed copy of src/stream/relay.js as
modules/services/openpost-radio/relay.js. Sync that deployment artifact,
set RGO_MUSIC_REVISION to its source commit, and apply with the configuration
repository's rebuild wizard (rebuild --vps). The unit executes the immutable
Nix-store copy, so a rollback restores both the sender and its configuration. Never
put stream keys in source, command arguments, or logs. The unit has a
128 MiB memory limit and a 50% CPU quota. status.json in
/var/lib/openpost-radio records output bytes, playback time and format;
measure actual cgroup CPU and memory separately from these limits.
The relay shifts the chunks' PCR, PTS and DTS onto one continuous clock,
including at the eight-hour programme wrap. Encoded media bytes stay unchanged.
test/relay-timestamps.test.js checks repeated wraps, synchronized audio/video,
and preservation of AAC frames. A healthy sender status is not playback proof:
verify a capture of YouTube's served audio against the programme after changes.
The normal programme completes dawn/day/dusk/night every 40 real
minutes, independent of a viewer's timezone. Seasons come round about every
five hours of travel. --day-minutes 0 retains real solar time for direct
rendering; the scripted --tour uses its own times.
The original audio library remains on the NAS at
/volume1/media/openpost-radio-music (moved out of the user home in October
2026 so Synology Drive stops indexing and versioning the radio chunks; a
symlink at the old home path bridges the relay's root-owned read jail, whose
ROOT should follow at the next admin pass). Generate compositions on a
workstation and publish with scripts/publish-nas-library.sh, which copies
finished audio before the catalog. Direct npm run stream playback still
supports a local library or an HTTP source with ten compatible tracks ahead,
atomic downloads, and protection for playing and fading tracks.
The generated, badge-free livestream thumbnail is
output/livestream-thumbnail.png.
Its full-resolution original and generation prompt are in assets/marketing/.
Code: see the repository licence. Strudel packages are AGPL-3.0. Font: Geist
Mono under the SIL Open Font License (assets/fonts).
Sounds come from Strudel's default libraries, fetched at render time: Tidal
drum machines, VCSL (CC0), the Salamander piano (CC-BY 3.0) and General MIDI
soundfonts. Check the sample licences before monetising the channel.