AquaSynth/docs/pink-trombone-dsl-parity-plan.md

Pink Trombone DSL Parity Plan

Objective

Cover Pink Trombone's useful pipeline with AquaSynth language constructs, then use Pink Trombone as an audio parity target. Do not build a PT clone with Aqua paint on it. The primitives must compose with ordinary Aqua voices, envelopes, LFOs, parameters, layers, learned speech controls, filters, and future morphologies.

Current Mechanism

tract currently lowers into a voice-local scalar control bundle plus a typed AcousticPortNetwork. The graph is the only vocal-tract audio authority. The old source/filter proxy and old tract waveguide backend are removed; legacy propagation names now fail fast instead of quietly rendering a fake voice. The remaining anatomy ownership is still too discrete:

• authored tract shapes now own physical length and continuous interpolation, and a tract voice may resample them to a chosen compiled grid;
• sections, nose_sections, and integer junction indices still leak lowering grid choices into the script surface;
• substeps is no longer a live lowering contract; acoustic length, propagation speed, and fractional delay own timing;
• topology changes can recompile, but morphology motion must remain runtime control data.

Invariants

• PT pressure must become Aqua DSL vocabulary, not a bespoke PT mode.
• A tract voice remains a voice. Tract primitives are treatments, sources, curves, junctions, and event lanes that can be combined with the rest of the patch graph.
• Diameter/area functions own continuous tract shape over normalized position, physical acoustic length, derived areas, reflection coefficients, and resampling. Tongue and constriction gestures deform that shape; they do not replace the owner.
• PT's 44 oral cells and 28 nasal cells are a reference discretization, not an Aqua morphology model.
• Reflection coefficients derive from adjacent areas. Caches and summaries may observe that derivation, but must not become independent truth.
• Wave clock derives from acoustic length, propagation speed, sample rate, and delay approximation. User-facing substeps must not own morphology timing.
• The typed acoustic graph owns delay-line propagation. Formant filters and tract proxies are not valid fallbacks for voice parity.
• Expressive parity comes before audio golf. Audio parity claims require rendered PT fixtures, Aqua renders, and log-mel cosine similarity.

Primitive Decomposition

1. tract_shape: reusable section diameter/area function.

   • Owns sampled diameters or areas over normalized continuous tract position.
   • Owns physical acoustic length through length_cm.
   • Emits interpolated diameters, resampled grids, areas, reflection coefficients, per-cell acoustic delay, and shape summaries.
   • Feeds tract voices and later morphology/gesture planning.

2. glottis: reusable excitation primitive.

   • Owns frequency, intensity, tenseness, aspiration, and reflection intent.
   • Can feed tract voices or other voice treatments.
   • Implemented as a named source primitive consumed by tract voices.

3. tract_injection: positioned noise/burst excitation.

   • Owns constriction position, diameter/opening, turbulence, and release transient behavior.
   • Can be driven by envelopes, LFOs, learned speech controls, or consonant gestures.
   • Implemented as named graph source ports consumed by tract voices and distributed over cell-centered source terminals by position and width.

4. nasal_junction: velum-controlled branch primitive.

   • Owns nasal opening and branch shape.
   • Derives the three-way junction coefficients from local areas.
   • Implemented as nasal_branch sugar over acoustic path terminals and a graph connection.

5. acoustic_graph: propagation primitive.

   • Owns right/left traveling-wave state, acoustic wave clock, fractional delay/loss approximation, and radiation.
   • Consumes area/reflection fields and source/injection events.
   • Graph lowering splits paths at typed terminals, emits bidirectional segment state, and scatters two-port and N-port junctions from live area and admittance.
   • Tract graph generation derives live per-terminal area from the tract shape plus tongue, constriction, and lip controls before scattering.

6. tract_motion: control-rate slew and obstruction history.

   • Owns diameter/constriction/velum slew rates and obstruction threshold.
   • Feeds smoothed target controls and release-transient detection.
   • Implemented as tract_motion; waveguide lowering uses it for control slew and stateful obstruction-opening burst pressure.

7. PT parity harness.

   • Renders fixed PT references for vowels, nasals, fricatives, closures, and moving tongue/constriction gestures.
   • Renders Aqua DSL reconstructions.
   • Scores log-mel cosine similarity plus targeted feature probes.
   • First fixture catalog exists in PinkTromboneParityFixtures.
   • PinkTromboneReferenceRenderer now renders deterministic test-only traveling-wave references for the fixture controls.
   • PinkTromboneLogMelParityTests writes reference/candidate WAVs and report artifacts under artifacts/parity/pink-trombone-logmel/.

First Cut

Implemented owners: tract_shape, glottis, tract_injection, nasal_branch, tract_motion, and graph-native propagation=graph. tract_shape owns section diameter/area fields and reflection derivation. glottis owns excitation quality. tract_injection owns positioned frication/burst pressure. The current Faust proxy consumes these primitives through shape summaries, reflection energy, glottal shaping, and injection pressure instead of silently inventing all of them inside one helper. The waveguide lowering consumes the derived oral reflection field as right/left section state equations.

The first log-mel rung is now real but not flattering. The first compileable Aqua proxy baseline against the deterministic PT-style renderer landed around cosine 0.51 open vowel, 0.54 front vowel, 0.39 nasal, 0.12 sibilant, and -0.18 closure-release. That negative closure score correctly exposed the wrong machine.

The old wg_loop ~ si.bus(...) backend is now commit history, not a live lowering. Graph output uses graph_loop ~ si.bus(...), physical segment lengths, and fractional delay. Generated tract graphs preserve declared morphology unless the author explicitly asks for another compiled grid.

The continuous morphology cut has started. tract_shape length_cm=... is now physical geometry, not just a list length. TractAreaFunction can interpolate diameter over normalized position, resample to an arbitrary grid, and derive per-cell acoustic delay. A tract voice may choose a different sections=... lowering grid without requiring the named shape's authoring samples to match. For a human-length 17 cm tract sampled at PT's 44 oral cells at 44.1 kHz, that delay is about 0.5 samples per cell. That makes PT's two tract updates per output sample look like a discretization strategy for a half-sample cell grid, not a reusable Aqua language concept.

The acoustic graph cut has also started. Aqua now has neutral model records for AcousticPath, AcousticTerminal, AcousticConnection, AcousticSourcePort, AcousticBranch, AcousticRadiationPort, WaveClockPolicy, and AcousticPortNetwork. Parser commands path, terminal, connect, source_port, branch, radiation_port, wave_clock, and acoustic_network can express a larynx, paired labial syrinx-like sources, a nasal/oral junction, or stranger source/topology combinations without adding a species mode. Existing PT-shaped commands populate those acoustic records as authoring aliases: source ports and radiation ports create typed terminals, while branch creates branch endpoint terminals plus a connection. acoustic and tract voices lower through the same compiled graph; missing or invalid graph topology renders silence with a warning rather than falling back to a pretend tract.

Graph ownership decision

The topology choice is now explicit: Aqua will use a path graph with typed terminals. Authoring sugar can hide some ceremony later, but the compiler-facing shape is geometry plus terminals plus connections. AcousticBranch is no longer the future topology law; it is a side-branch shorthand over the graph.

The first graph lowering cut now compiles typed graphs into bidirectional Faust segment state. It splits each path at terminal positions, orders connection ports deterministically, scatters connection groups with area-weighted pressure continuity, injects source terminals, and sums radiation terminals. Faust validation passes on a three-path trachea/oral/nasal graph.

The graph compiler now also honors wave_clock delay strategy. Segment delay comes from physical segment length, propagation speed, and ma.SR, then lowers through unit-grid, half-sample, linear fractional, Lagrange, or Thiran/allpass Faust delay primitives. Co-located terminals on one path now aggregate into a single graph node, so a source can share an anatomical point with a junction or boundary without disappearing.

The remaining implementation pressure is now narrower: broaden connection laws with sharper physical semantics and keep the log-mel parity loop on the graph renderer. PT-shaped tract authoring now uses propagation=graph for accepted parity. The old waveguide lane is not a baseline, pressure lane, or fallback.

Authority map for the required rebuild:

• Owner: the acoustic network lowering owns an ordered scattering graph, not Voice.Tract, not the removed tract waveguide, and not a response proxy.
• Inputs: continuous paths, typed terminals, connections, source/radiation ports, branch sugar, wave-clock policy, and live parameter bindings.
• Outputs: Faust state for delay-line segments, source injection points, junction scattering, radiation, and diagnostics.
• Derived state: tract/nasal_branch are larynx/PT-shaped authoring conveniences over the same acoustic graph.
• Forbidden writers: Voice.Tract, legacy propagation modes, and response proxies must not produce alternate vocal audio; branch/radiation/source controls must not be repaired by a separate playground model.
• Shared paths: direct DSL authoring, PT compatibility commands, future syrinx/reed/alien voices, and training/playground parameter bindings must all commit through the same acoustic graph.
• Deletion line: do not expand the proxy or revive the old waveguide into a second fake synthesizer. AcousticTerminal/AcousticConnection records own graph lowering; branch can remain only as shorthand that emits those records.

Exact Pink Trombone timing still needs pressure, but the next backend cut is no longer "add more substeps." The graph renderer now has fractional-delay clock semantics and a tract propagation=graph entry point; the remaining parity question is whether graph-authored PT compatibility can beat the current weak Voice.Tract baseline.

The parity report now renders the graph lane only. The latest static fixture evidence is useful instead of decorative: open/front/nasal/ma/sibilant/closure cosine is 0.6067/0.6573/0.6871/0.6476/0.4502/0.6168. The useful levers so far are source semantics, radiation impedance, terminal-specific admittance, frication injection, cutting stale first-frame morphology from generated graph paths, and live path-area control at graph sample points.

The nasal branch pressure exposed the next graph-law issue. The graph now uses terminal-specific admittance in connection pressure instead of collapsing every co-located terminal into one node area, and branch output uses a floored area-ratio pressure normalization so weak side ports no longer receive full junction pressure. The generated tract graph also stopped treating the first nasal tube diameter as the velum aperture; velum now owns branch entrance area and nasal aperture, while connection coupling remains a structural admittance limit. Two tempting compensators were cut: raw per-port outgoing coupling made supposedly oral fixtures nearly silent, and lip/beak radiation gain did not restore oral vowel RMS. The internal-node scatter fix made unconnected injection/source terminals transmit pressure between adjacent segments instead of acting as dead boundaries; latest utterance cosine is mama 0.3365, papa 0.8264, and thrombosis 0.3874.

Live graph morphology now has a path-level owner. AcousticPath.AreaControl turns tongue, constriction, and lip controls into live diameter/area expressions at generated graph terminals, and generated tract propagation=graph paths insert neutral area terminals along the oral tract so those live areas actually scatter waves. This is still sampled continuous morphology, not a resurrection of PT's fixed cell module.

Frication also had a low-level ownership bug: generated graph TurbulenceJet ports were using burst as source pressure, so sustained sibilants were choked by a transient control. They now use max(turbulence, burst) as pressure while keeping turbulence as the noise amount. This only nudges the current sibilant fixture (0.2302 cosine), but the primitive semantics are cleaner: turbulence owns continuous noisy pressure, burst owns transient pressure.

The graph source law now separates sustained aperture noise from closure release. AcousticSourcePort.Transient owns release energy, while sustained TurbulenceJet noise is gated by the intersection of the old upper constriction cap and a narrow-open aperture window. Closed constrictions no longer hiss just because they are closed. Latest accepted utterance graph smoke is mama 0.3339, papa 0.8320, and thrombosis 0.3957 under artifacts/parity/pink-trombone-utterance-logmel/20260527T110608640.

Moving source position is now graph-native too. AcousticSourcePositionControl weights fixed source terminals from a live index/width/scale expression, and generated tract propagation=graph injection emits an oral source-port bank instead of one stale terminal. This keeps the graph topology Faust-friendly while letting frication and release energy follow the articulator. Latest accepted utterance graph smoke is mama 0.3418, papa 0.8318, and thrombosis 0.4521 under artifacts/parity/pink-trombone-utterance-logmel/20260527T111349174.

The first listening-led anti-buzz cut changed the glottal source, not the tract topology. A direct LF coefficient expression improved some cosine scores but overdrove the recursive graph and broke the static open-vowel fixture, so it was cut. The kept graph source adds a mild cubic waveshaping component to the existing glottal primitive. Latest accepted utterance graph smoke is mama 0.4822, papa 0.8602, and thrombosis 0.5202 under artifacts/parity/pink-trombone-utterance-logmel/20260527T114901634. The band probe still shows weak 500-1000 Hz vowel-body energy, so this is not the final answer to the smartphone-vibration complaint. It only proves source excitation was part of the failure; tract scattering/radiation remains the live pressure.

The next successful anti-buzz cut moved ordinary same-path interior nodes from transparent pressure averaging to signed two-port area-discontinuity scattering. The first version used a midpoint admittance pressure law and collapsed mama; the kept version uses adjacent segment areas with the reflection sign that matches the graph's right/left state convention. Latest accepted utterance graph smoke is mama 0.3262, papa 0.8567, and thrombosis 0.5762 under artifacts/parity/pink-trombone-utterance-logmel/20260527T120741887. The band probe now shows papa with roughly half its energy in the 500-1000 Hz body band instead of mostly 80-200 Hz vibration. mama still lacks enough 500-1000 Hz vowel body, so nasal/oral coupling and radiation remain live pressure.

The next radiation cut moved aperture out of the output-only lane. Lip/beak opening now blends graph boundary reflection between a near-closed termination and the declared open-end reflection. Radiation reads boundary flow (incoming - outgoing) rather than raw incoming pressure, then applies a stronger radiation slope. Latest accepted utterance smoke is mama 0.5650, papa 0.8613, and thrombosis 0.4347 under artifacts/parity/pink-trombone-utterance-logmel/20260527T122757617; latest static smoke is open/front/nasal/ma/sibilant/closure 0.6088/0.6826/0.6468/0.6153/0.4158/0.6198 under artifacts/parity/pink-trombone-logmel/20260527T123141625. This is a real authority fix, not the final mouth. The remaining articulation failure points at pressure storage/release around severe constrictions and better nasal/oral coupling, not another output color patch.

A small glottal-brightness pass then added higher modal harmonics to the reusable glottal source rather than adding a PT-only exciter. The post-flow radiation normalization now uses 4.0, a middle point that improves utterance RMS without pushing static closure into the soft-clip guardrail. Latest accepted utterance smoke is mama 0.5712, papa 0.8659, and thrombosis 0.4379 under artifacts/parity/pink-trombone-utterance-logmel/20260527T124230804; static smoke remains passing under artifacts/parity/pink-trombone-logmel/20260527T124311305. This helps the voiced fixtures a little, but it is not the missing articulation mechanism.

Log-mel cosine is no longer allowed to impersonate a speech-parity verdict. It rewards spectral overlap even when the candidate is mostly silence, filtered motor buzz, or a few breathy/squeaky bursts. AudioComparison now includes an AudioArticulationComparison with envelope cosine, active-frame ratio, silence mismatch, envelope/spectral flux ratios, motor-band ratio, speech-band ratio, and an articulation score. The utterance report prints a verdict. Latest run: mama, papa, and thrombosis are all not-accepted-articulation under artifacts/parity/pink-trombone-utterance-logmel/20260527T130026316, despite their old cosine values. That is the harness telling the truth again.

The next physically grounded cut fixed graph tract timing. Segment delay is no longer clamped to a full sample before wave-clock lowering; the wave-clock strategy owns the minimum legal delay. Generated tract propagation=graph patches now use HalfSampleGrid, matching the half-sample Kelly-Lochbaum regime that PT approximates by updating the tract twice per audio sample. This removed an accidental overlong-tract crutch and improved several witnesses, especially thrombosis motor-band ratio. Latest accepted utterance smoke is mama 0.5982 / articulation 0.1952, papa 0.8615 / 0.1979, and thrombosis 0.5217 / 0.2640 under artifacts/parity/pink-trombone-utterance-logmel/20260527T133539554. The voice is still not accepted; the remaining failure is body/radiation and pressure-release behavior, not tract length.

Cut Line

The current Voice.Tract scalar proxy may survive only as an adapter over real tract primitives. If a future change adds PT behavior that cannot be described as a reusable Aqua primitive, cut it or keep it in the reference harness instead of promoting it to the public DSL.