Transcription

Live mode (Mode 0): the node drains the connected Audio ring continuously on a background task, converts to the model's format and streams it into Apple's on-device SpeechAnalyzer. Instant volatile guesses appear first (marked final:false, no reliable times, confidence −1) and are later replaced in place by finalized words carrying word-level start/duration from the model — when the finalization changed the text, Corrected pulses, so displays can animate the fix. Live captions inherently trail speech by 1–3 s at finalization; the per-word age field (wall-clock seconds since the word appeared) is what kinetic displays should animate on.

Timeline mode (Mode 1): cable a player's Info output into Source — one file pick in the player feeds playback AND transcription (the manual File path is the fallback for player-less patches) — and the node pre-transcribes the whole file once, faster than realtime. The Time input (cable it from the player's Time output) then serves Text/Word/Segments exactly as of that playhead — zero lag, scrub-safe in both directions, loops follow. In this mode age = Time − start (audio-clock), and NEGATIVE age means the word is still coming — kinetics can pre-stage it.

Speaking comes from the SpeechDetector voice-activity module riding the same analyzer. The Segments output is the kinetic-text contract: an array of dicts {text, index, start, duration, end, confidence, final, sentence, age}, indices absolute and stable.

State is volatile (project load starts empty); Reset clears on a rising edge; locale/censor/source changes rebuild the session. The engine feeds from a private accumulator, so every Audio Input Quality preset works — the ring only has to survive one 40 ms poll.

Audio (structure) — any Audio port. For pristine results transcribe the SOURCE directly (Audio/Video Player cable), not a microphone listening to speakers.

Mode (int dropdown) — Live / Timeline; ports follow the mode. Source (structure) — a player's Info output: the file connection of record in Timeline mode. File (string) — manual fallback path (hidden while Source is cabled). Time (float) — Timeline playhead in seconds.

Locale (string, default en-US) — inspector dropdown lists the 30 supported locales; ⇣ marks models not yet downloaded (they install once, system-wide). Unsupported locales report honestly in Info. Latvian is not currently supported by Apple's model.

Censor — etiquette replacements (f··· style). Max Words — rolling transcript window (hard cap 5000); transcript LENGTH is governed here, never by audio buffer sizes. Reset — rising edge clears.

Outputs: Text (finalized), Live Text (+volatile tail), Sentence, Word, Segments (the contract), pulses New Word/New Sentence/Corrected, Speaking (VAD level), WPM (10 s window), Confidence (last final result), Pause (seconds of silence), Info (status/locale/words/ring warnings).

• Live captions — Audio Input → Transcription → Live Text → UI Text: on-stage captions with zero cloud.
• Karaoke — Audio Player Info → Source, Time → Time (Timeline mode); Word → UI Text: one file pick, and the current word lands exactly on the beat of the voice.
• Voice-triggered visuals — Word → String Compare ("storm") → Magic Storm trigger: say the magic word.
• Silence dramaturgy — Pause → Smooth → Drone brightness: the longer nobody speaks, the heavier the room gets.
• Speech-rate energy — WPM → particle turbulence: fast talkers whip the swarm.
• Subtitles for delivery — Timeline mode + inspector Export SRT/VTT: broadcast-format subtitle files (42 chars, 2 lines) from any media file.

• One-frame pulses (New Word) chain perfectly into Hold Trigger for visible flashes.
• Segments' age is the animation clock: fade words in by age, ghost the final:false tail, snap solid on finalization.
• Two nodes, two locales, one mic = simultaneous bilingual captions.
• Brand names and invented words will be misheard — the model formats numerals ("42", "$5") but knows no custom vocabulary yet.
• Mic quality honesty: room echo and speaker playback degrade recognition heavily; cable the digital source whenever one exists.