MIDI 2.0-native singing instruments for Fountain Coach.
Status, 24 October 2026: the native Swift synthesizer now renders three independent, SHA-256-verified WAV takes from one five-second synthetic score: plain, articulated, and expressive. The articulation mode prolongs /m/ and /n/ nasal onsets and shapes their transition to vowels; the expressive mode adds typed MIDI 2.0 pitch bend and breath, plus optional vibrato. Four notes retain three syllable articulations across melisma. A separate macOS PCM output adapter works without CoreMIDI, and full authored WAV outputs were independently reproduced byte-for-byte outside the Mac. No proprietary voice bank, no alternate MIDI plane. Human listening, external MIDI transport, hard-real-time limits and publication remain separate pending gates.
A voice is an instrument, a syllable is a timed phonetic state, and a performance is an executable score. sing makes text, phonemes, notes, expression, performer identity, and verifiable sound independently addressable. Our own glottal-excitation/formant synthesis is the implementation path. Alter/Ego and Wolfgang Palm's PPG Phonem remain historical, black-box references only; neither proprietary implementation is part of this source or required to build and render.
- Satzfunk (private source authority) is the manuscript-addressed listening, editorial approval and publication context: First Reading → First Hearing.
- Reframe owns source text, selected passage, edition, identity and reader/writer decisions.
- Teatro, ScoreKit and AudioTalk express scenes, temporal notation, performance intent.
- midi2 is the typed Universal MIDI Packet and instrument-transport foundation; midi2sampler contributes sample DSP, not a completed singer.
- sing owns phonetic alignment, voice identity/capability contracts, renderer-neutral vocal performance, comparison and instrument tests. No duplicate editorial coordinator, MIDI transport or generic DAW.
- instruments.fountain.coach may catalogue admitted instrument identities; it is a supporting reference domain, not a new public editorial front.
A source passage and its hash are not a voice bank. A voice bank and its licence are not a score. An executable score and its renderer are not a published First Hearing. Keeping these separate will let one score be performed by an independent procedural singer, a consented sample voice, or an explicitly qualified future neural model. Historic vendor compatibility is optional research, not a product gate.
7VST / Seven Against Thebes is the documented historical specimen: a 2017 Alter/Ego/Bidule/ARIA experiment and privately restored audio/text evidence. The author also reports extensive use of PPG Phonem for iPad in the historical Bandcamp work; particular track assignments have not yet been independently established. New renditions are new editions, never silently labelled as the historical production. See 7VST lineage.
Satzfunk offspring may be independently versioned voices, choirs, pronunciation models, scores, or small autonomous performance agents with explicit parentage, controlled execution, provenance and rights. “Babies” means independent artistic instruments, not unbounded code replication. See offspring contract.
- 23 October: native voice first and verified boundaries
- 23 October: typed MIDI 2.0 UMP rehearsal evidence
- 23 October: real-time voice callback and macOS audio-output witness
- 24 October: extended native expressive audition and physical output witness
- 24 October: three-take listening comparison and verified checksums
- 24 October: actual Ear worklet chooses the articulated development baseline
- 24 October: speech-to-singing private experimental renderer, Ear pitch evidence, and source-blind lexical gate/voice-envelope findings — MIDI2 notes verified, on-device recognizer does NOT recover all lyrics, no accepted naturalistic voice or OpenAI capture; source-only development.
- 25 October: Ear/Sing teaching-tradition design (Italian, German, Spanish, tabla and Hindustani vocal study) — text-only curriculum fixture, no trained tutor or qualified recognition.
- Architecture and instrument boundaries
- Code audit: Fountain Coach + Contexter and immutable upstream Git pins
- Alter/Ego compatibility reference
- Wolfgang Palm's PPG Phonem for iPad: resonator voice and historical Bandcamp testimony
- 7VST historical source and new score
- Satzfunk family / offspring
- Executable acceptance scenario and roadmap
- Publishing, consent and rights
- Provisional performance JSON schema, short synthetic phrase, and five-second sustained melisma
Current milestone: all three synthetic takes were passed to Teatro's actual Ear AudioWorklet, plus a separately scoped test-only Ear spectral-analysis worklet in a real Chrome Web Audio graph. The observed nasal onset attenuation and steady-vowel parity select the articulated take as the next phonetic-development baseline; the expressive take remains a distinct performance option. Ear does not yet recognize lyrics, and its shipped note 48 output is a fixed placeholder, not measured pitch. The acceptance gap is an independently witnessed phoneme recognizer owned by Ear, not a request for the author to grade engineering test recordings. Editorial approval of a published performance remains author-owned. See Ear's real-worklet three-take witness. No CoreMIDI, vendor voice, private sound export or public release was added.
swift test
swift run -c release sing-render examples/three-syllables.json /tmp/sing-la-mi-na.wav
swift run -c release sing-render examples/three-syllables.json /tmp/sing-control.wav --no-notes-control
swift run -c release sing-midi2-rehearsal examples/three-syllables.json /tmp/sing-ump.wav
swift run -c release sing-midi2-rehearsal examples/three-syllables.json /tmp/sing-pressure.wav --pressure
swift run -c release sing-midi2-rehearsal examples/three-syllables.json /tmp/sing-realtime.wav --realtime
swift run -c release sing-midi2-rehearsal examples/sustained-melisma.json /tmp/sing-plain.wav --realtime
swift run -c release sing-midi2-rehearsal examples/sustained-melisma.json /tmp/sing-articulated.wav --realtime --articulated
swift run -c release sing-midi2-rehearsal examples/sustained-melisma.json /tmp/sing-expressive.wav --realtime --articulated --expressive --vibrato
swift run sing-macos-output examples/three-syllables.json --probe
# plays quietly on the machine's default output only when explicitly requested:
swift run sing-macos-output examples/sustained-melisma.json --play --articulated --expressive --vibrato
python3 -m unittest discover -s tests -v
python3 tools/check_phrase.py examples/three-syllables.json
python3 tools/check_phrase.py examples/sustained-melisma.jsonThe Swift tests measure deterministic PCM, typed MIDI 2.0 Note On/Off, Polyphonic Pressure, Pitch Bend and Breath CC2 events, malformed/reserved-field rejection, fundamental-frequency shifts, noise-vs-loudness distinction, buffer-size invariance, sustained melisma, optional vibrato, quiet release and CoreMIDI exclusion. They also verify WAV output, reject unknown phonemes and require silence for the no-note control. The Python validator checks phrase/note alignment and rejects false authored rendering or release states, with optional full JSON Schema validation when the jsonschema package is installed. These tests verify sample production, not perceived intelligibility, human consent, musicality or physical MIDI delivery. No GitHub CI is configured.
Audited code ≠ built package ≠ integrated renderer ≠ audible output ≠ authorized release. Every milestone must name its evidence. Public links to private repositories resolve only for authorized readers. Do not run GitHub Actions or publish a site as an incidental side effect. Git is source authority; deployment is separately authorized and verified.
Created 2026-10-21. © Fountain Coach contributors; rights of source materials and dependencies remain with their respective owners. No blanket code/media licence is granted by repository visibility.