Audio engine

Signal path

Strudel source → official transpiler → official pattern scheduler
  → SuperDough voices and per-orbit effects
  → studio orbit gain / stereo pan / analyser
  → SuperDough destination gain (master volume)
      ├─ browser audio destination
      ├─ master FFT / waveform / stereo analysis
      └─ PCM capture AudioWorklet → WAV blob

The implementation does not substitute a small oscillator engine for Strudel. Mini notation, scheduling, synths, samples, wavetables, envelopes and effects remain upstream functionality. Studio orbit controls sit after the orbit effects and are separate from each pattern’s own gain/pan parameters.

Startup and stopping

AudioContext creation is lazy in upstream code. Play/Evaluate resumes it after user activation and loads official worklets. Safari may require a second click directly in the editor frame. Loading a shared score never calls evaluation or autoplay.

Hush immediately gates the master output, stops the scheduler, finishes any recording, and resets the official SuperDough output/orbit graph before suspending the context. Suspending alone would freeze scheduled voices and effect tails, which could briefly resume when another file starts. Resetting disconnects those old voices while retaining registered samples, decoded sample caches and mixer settings. The next evaluation waits for an in-progress stop and opens the master gate only after evaluation. Already suspended contexts do not receive a redundant suspend request, avoiding a WebKit promise that can remain pending before the first audio gesture.

The browser regression records PCM after switching from a sustained, reverberant score to a different file with an initial rest. That rest must remain silent before the new score begins. Ordinary live updates continue to use Strudel's scheduler without resetting the audio graph.

Live updates and anonymous tracks

Upstream Strudel creates a new internal track ID each time $: is evaluated with its shouldHush argument set to false. The original block-evaluation integration retained those IDs, so repeated Mod+Enter added copies of the same pattern and increased the output level.

The runtime now uses CodeMirror's JavaScript syntax tree to identify anonymous labels, including S$:. For these scores, selection/block/line updates evaluate the current whole file with Strudel's normal pattern replacement enabled. This clears the pattern registry, preserves all current tracks, and leaves the audio graph and scheduler running. Source code is unchanged. A console message explains the whole-score update. Named labels such as drums: can still update one block while other named tracks keep playing. Switching files also replaces the score on the first evaluation.

Background music

Open Preferences → Audio / Background Playback → Allow background music. The setting is on by default and persists in this browser. Turn it off to hush when the studio tab becomes hidden or the browser is minimized. That also finishes an active recording; returning to the tab does not restart it. The command palette includes Audio: Toggle Background Music.

Background mode keeps the current Strudel scheduler running. An AudioWorklet supplies the official scheduler's setInterval / clearInterval hooks with wakeups from the audio clock, so it does not depend on throttled window timers or animation frames. It does not replace Strudel's pattern scheduler. Orbit routing continues on scheduler wakeups; visual rendering and analyser messages stop while hidden. Hush disconnects the clock node and suspends the audio context.

Where supported, navigator.audioSession.type = "playback" declares music playback to Safari. The host also supplies Media Session title/artist metadata and system Pause/Stop handlers. These APIs are feature-detected; no browser policy is bypassed. If the browser interrupts an active context, returning to the studio attempts to resume that same context, with an Enable Audio button if a gesture is required. It never reevaluates source to resume playback.

The browser and operating system retain control. Safari/iOS versions, battery-saving settings, competing audio, device locking and sleeping can suspend the page or audio thread. The checkbox cannot guarantee uninterrupted playback in those conditions. Play always requires an initial user action; closing the tab stops playback.

References: Audio Session, page visibility policies, and WebKit background-audio issue.

Metering and visuals

A 1024-point master analyser supplies decimated waveform and frequency arrays at up to 20 updates per second. Independent stereo and orbit analysers provide RMS levels; the stereo waveform pair drives the phase plot. The spectrogram retains successive measured FFT columns. The UI displays RMS and peak dBFS and flags peaks near full scale. It does not claim to be a calibrated loudness meter or a mastering limiter. There is no always-on nonlinear limiter changing normal Strudel sound. Keep headroom in stacked scores.

Pattern event views query the actual current pattern around the scheduler’s current cycle. Their data includes note/sample, orbit, gains, effect values and durations. Inline Strudel visuals and source-token highlighting remain official implementations.

Canvas rendering is separate from audio. It caps display resolution, supports low/balanced/high/auto quality, reduces frame rate when rendering is expensive, honors reduced-motion settings and skips decorative work in hidden tabs. Browser scheduling and device limits still affect very large scores.

The separate Visualizer uses this same measured signal for nine generative scenes, four overlay layers and automatic scene changes aligned to Strudel cycles. It adds no audio processing or independent FFT.

Recording

public/pcm-worklet.js captures actual master output in stereo float blocks. The host receives an uncompressed IEEE float32 WAV, preserving the captured PCM values. Recording can pause/resume and stops at the earlier configured time or byte budget. The raw take can be downloaded before any upload.

The server inspects recording uploads and uses FFmpeg to export:

• WAV: 24-bit PCM.
• MP3: libmp3lame, 320 kbps, ID3v2.3 metadata.
• M4A: AAC, 256 kbps, fast-start container.

The title, artist, album and comment come from bounded metadata. A public project URL is included only for public/unlisted projects. A generated PNG cover is attached to MP3/M4A as album artwork. The SVG master stays separate. Conversion is explicitly requested; no server audio process runs continuously.

Sample Lab

Sample Lab has its own trusted browser audio context for decoding, auditioning and microphone recording. PCM operations are applied to selected regions, support undo snapshots and leave channel alignment intact. Normalize uses a shared stereo peak. Resampling uses linear interpolation; playback rate changes pitch and speed together. It is not advertised as high-quality independent time stretching.

Diagnostics and limitations

Sample rate and base latency are reported from AudioContext. Values vary by device/browser. MIDI output and clock require supported ports; permission comes from the Controls UI. MIDI/serial/OSC hardware timing must be measured on actual equipment. No headless browser test can verify an external synthesizer or microphone’s physical signal path.