27 KiB
Acoustic Graph DSP Audit
This audit asks the blunt question for each live graph-voice component:
- What is this simulating?
- Does the implementation still make sense?
- If not, what authority should own the correction?
The answer is mixed. The graph is no longer pure rubble, but several pieces are still standing because they recently helped a metric rather than because they have a clean physical contract.
Pipeline Map
PatchScript builds neutral acoustic records:
AcousticPathowns tract or branch length, rest area, propagation speed, loss, and optional live area controls.AcousticTerminalmarks topology, source, branch, and radiation positions on paths.AcousticConnectionconnects terminals into scatter groups.AcousticSourcePortowns glottal, turbulence, reed, labial, click, or synthetic excitation.AcousticRadiationPortowns output boundary behavior.AcousticPortNetworknames which paths, terminals, connections, sources, radiation ports, and wave clock form one graph.
FaustExport lowers that graph into bidirectional delay state:
- Terminals become graph nodes.
- Adjacent nodes on each path become bidirectional segments.
- Each segment stores right-going and left-going traveling waves.
- Nodes scatter incoming waves into outgoing waves.
- Source ports add energy at node context.
- Radiation ports read boundary flow.
- Segment delays propagate outgoing waves into the next sample state.
That is the right high-level architecture. The weak parts are inside the lowering laws.
2026-05-28 Legacy Vocal Renderer Cut
Owner: AcousticPortNetwork and typed graph lowering own vocal acoustics.
Inputs: paths, terminals, connections, source ports, radiation ports, wave clock, and ordinary Aqua parameter bindings.
Outputs: one Faust graph loop plus named source, scattering, segment delay, and radiation signals.
Derived state: tract, glottis, tract_injection, and nasal_branch are
authoring conveniences that populate the acoustic graph records.
Forbidden writers: the old Voice.Tract resonator, the old Voice.Tract
waveguide backend, and the acoustic response proxy no longer produce vocal
audio. Legacy propagation=resonator, propagation=waveguide, and related
aliases now fail at parse time. Invalid graph topology emits silence with a
warning instead of inventing a fallback body.
Deletion line: do not restore a fake vocal renderer to make a test pass. If the graph cannot express a vocal behavior, add or fix graph semantics.
2026-05-28 Source Load And Contact Cut
Owner: graph source ports own source impedance/load interaction; graph contact terminals own closure pressure storage and release at constrictions.
Inputs: source pressure/tension/opening/noise/balance/impedance, node incident pressure, contact terminal area, and upstream/downstream traveling waves.
Outputs: load-aware source flow terms, named _load_pressure locals for
non-turbulence sources, and named graph_contact_* closure/reservoir/release
signals for contact terminals.
Derived state: generated tract area terminals are topology junctions that derive area from the live tract path geometry; generated contact terminals are a sparse overlay that owns closure pressure storage and release. Area samples are no longer contact owners just because they define tube geometry.
Forbidden writers: plosive release is not allowed to live only inside a TurbulenceJet source expression. A source-local burst can remain as excitation, but graph contact state now has the authority to store and release pressure at severe constrictions.
Syrinx proof: source_port kind=syrinx is still the generic labial source
kind, now load-aware through the same source impedance path. The checked-in
patches/advanced/bird-syrinx.aqua patch uses two labial source ports, two
bronchial paths, one tracheal path, a three-terminal area-scattering merge,
and a beak radiation port. No species module was added.
2026-05-29 Sparse Contact Owner
Owner: generated tract authoring owns a sparse contact overlay; graph path
lowering owns contact release injection through ordinary two-port scattering.
Inputs: full area-junction grid for tube geometry, up to four generated contact terminals spread across the tract interior, local terminal area, node incident pressure, and upstream/downstream traveling waves.
Outputs: named graph_contact_* closure/reservoir/release signals only at the
sparse contact owners, while every generated voices_0_area_* terminal remains
a Junction for geometry/scattering.
Derived state: voices_0_contact_* is no longer a second morphology grid. It is
a small collision/pressure-storage overlay derived from tract section count.
Forbidden writers: ordinary area terminals must not decide plosive contact state. Inline contact occlusion in the scattering equation is not shipped; the first attempt still overflowed the native Faust debug compiler at eight contact owners, so contact count is capped at four until occlusion can be expressed as a smaller compiled primitive or a lower-pressure diagnostic path.
Evidence: opt-in thrombosis probe
artifacts/parity/pink-trombone-graph-thrombosis-probes/20260529T122217729
passes with lip-end contact closure peak 0.936876, contact output peak
0.547741, lip boundary load 0.516196, lip flow 0.585704, and modal source
out 1.235997. Accepted utterance diagnostic
artifacts/parity/pink-trombone-utterance-logmel/20260529T122409541 still fails
articulation: mama/papa/thrombosis cosine 0.8302/0.7865/0.3218, with
thrombosis RMS ratio 0.0770, active ratio 0.3151, and silence mismatch
0.7205.
2026-05-29 Compact Generated Tract Graph
Owner: generated tract authoring owns the translation from high-resolution
tract morphology into a Faust-sized graph. VocalTract.Sections still owns
morphology and index semantics; generated acoustic terminals/source ports own
only the compiled approximation.
Inputs: full tract area function, tract section count, live tongue/constriction and lip controls, nasal branch position, and injection position/width.
Outputs: at most ten generated oral area junctions, at most eight generated injection source ports, and at most four generated contact terminals. Source position weighting is widened to the compact source spacing so a moving constriction remains represented without one source port per tract cell.
Derived state: generated voices_0_area_* and voices_0_inj_* records are no
longer a one-record-per-section discretization. They are backend lowering
samples of the continuous tract owner.
Forbidden writers: the Faust backend must not be forced to compile the full authoring grid. If a future change needs more articulatory precision, it should improve the continuous area/source/contact heuristics or add a named compact primitive, not restore a 44-node/43-source generated graph by default.
Evidence: latest opt-in thrombosis debug DSP
artifacts/parity/pink-trombone-graph-thrombosis-probes/20260529T125048965
dropped from 1652 lines to 532: source-related graph lines 687 -> 162,
segment-area lines 211 -> 56, next-state lines 89 -> 27, and area-reflection
lines 74 -> 12. The opt-in native graph probe passed in about 1m15s.
Accepted utterance diagnostic
artifacts/parity/pink-trombone-utterance-logmel/20260529T125227926 passed in
about 1m33s; mama/papa/thrombosis cosine is 0.8324/0.7730/0.3176.
Thrombosis RMS improved to 0.1715, but active ratio fell to 0.1598 and
silence mismatch rose to 0.7811, so articulation remains the live failure.
Follow-up: source placement now uses a raised-cosine kernel instead of a hard
triangular weight over the compact source bank. It keeps the same graph size
while making moving injection ownership continuous enough for DSP. Artifact
artifacts/parity/pink-trombone-utterance-logmel/20260529T140038717 nudges
thrombosis active ratio to 0.1826, silence mismatch to 0.7643, and
speech-band ratio to 0.2704; still not accepted, but the improvement came
without restoring discrete source density.
Follow-up: generated lip radiation now mirrors live tract lip opening into
/acoustic/radiation/*/opening. Before this, path-end area followed the mouth
gesture but the radiation aperture stayed at its default, splitting ownership
between geometry and output. Artifact
artifacts/parity/pink-trombone-utterance-logmel/20260529T141822464 improves
papa RMS ratio 0.3242 -> 0.8369, cosine 0.7725 -> 0.7996, and motor-band
ratio 0.5238 -> 0.0555, which matches the listening complaint that plosives
were drum hits. mama remains wrong and thrombosis remains mostly absent, so
the next vowel cut is tract/radiation color and vowel body, not more plosive
gain.
Component Audit
Topology Construction
What it simulates: fixed graph discretization of acoustic paths, with topology terminals defining scattering locations.
Current state: mostly coherent. Interior source terminals no longer create topology nodes; they attach to existing graph nodes. That fixed a real authority leak.
Smells:
- Interior sources attach to the nearest node, not an interpolated point inside a segment. Generated tract source banks compensate by creating many fixed source terminals. That is expressive but heavy.
- A node can effectively belong to one connection through
nodeConnection. More complex graph topologies may need a first-class multi-connection node model instead of assuming one branch connection owns the node. terminalConnectionis built but not used. That is small rot.
Verdict: keep the model, but source placement wants an interpolated injection primitive before this becomes a clean alien-tract substrate.
Area And Geometry
What it simulates: time-varying tract cross-sectional area derived from rest diameter plus tongue, constriction, and lip controls.
Current state: useful but simplistic. Diameter is squared into area, which is the correct broad convention. Generated lip opening now owns both path-end geometry and radiation aperture.
Smells:
- Tongue/lip shaping is a hand-authored Gaussian blend, and constriction is a min-like clamp. This is a serviceable control surface, not a principled articulatory model.
- Graph propagation does not use
tract_motionslew. Utterance fixtures smooth controls externally, but generic graph voices can move morphology abruptly. - Node area sums co-located non-source terminal areas. That is right for some admittance joins and suspicious for ordinary labels sharing a location.
Verdict: acceptable as a compact morphology surface, but motion/smoothing must move into graph area ownership instead of living only in fixture curves or the old tract proxy.
Wave Clock And Delay
What it simulates: travel time through each path segment.
Current state: conceptually correct. Segment delay follows physical length, speed of sound, and the selected fractional-delay strategy.
Smells:
- Audio-rate delay modulation from live morphology can hit Faust fractional delay primitives directly. That is flexible, but the graph has no explicit policy for smoothing delay changes or preserving passivity during fast topology-scale motion.
- Minimum-delay policy is hidden in
WaveClockMinimumDelay; it deserves clearer documentation per strategy.
Verdict: keep fractional delay. Add explicit delay-smoothing/passivity policy before trusting violent morphology animation.
Two-Port Path Scattering
What it simulates: Kelly-Lochbaum area-discontinuity scattering between two adjacent tube sections.
Current state: coherent. The graph now uses the PT/Kelly-Lochbaum traveling-wave convention:
r = (A_left - A_right) / (A_left + A_right).
Smells:
- Source injection is added equally to both outgoing directions at a two-port node. That is plausible as a symmetric pressure/flow source only when the source is actually centered at the discontinuity. It is not yet a general source impedance law.
Verdict: keep the scatter law. Revisit source injection weighting separately.
Three-Port Branch Scattering
What it simulates: PT-style oral/nasal side branch scattering.
Current state: replaced. The old recognized three-port branch path has been collapsed into the same general N-port scatter law used for every area scattering connection.
Smells:
- Source injection is added per port inside the branch law. That needs a clearer impedance story if sources can live at branch nodes.
- The new general law still uses scalar terminal area/admittance, not a frequency-dependent junction impedance.
Verdict: keep the generalization. It removes the PT-shaped topology shortcut, but source impedance at junction nodes still needs a named model.
Generic Connection Fallback
What it simulates: arbitrary connection scattering when a connection is not the recognized three-port branch.
Current state: cut. The lowering no longer emits
graph_connection_pressure_*, no longer uses the handmade
0.6 + 0.4 * sqrt(...) port-admittance blend, and no longer keeps a
shape-specific three-port branch path. Area-scattering connections now emit one
passive N-port traveling-wave law:
r_p = (2 * A_p - sum(A)) / sum(A)
out_p = r_p * in_p + (1 + r_p) * sum(other inputs)
Connection coupling now crossfades between that scattered wave and the local incoming wave. Bypass remains explicitly bypass.
Smells:
- Port area selection is still simple: ordinary isolated junction labels use adjacent segment area, while branch/radiation/source terminals use their declared terminal area. That is coherent enough for the present graph, but it should be documented as terminal admittance ownership before more exotic junctions rely on it.
- Source injection is still added as node pressure energy after scattering, divided by local node port count. That is not a full source impedance law.
Verdict: the highest-priority fallback Jenga is gone. Keep passivity probes on this law while cutting source impedance and closure storage next.
Glottal Source
What it simulates: voiced excitation at a source boundary.
Current state: better after local tanh normalization. Before that, the source
could dominate the tract as an overdriven buzzer.
Smells:
- The waveform is still a hand-shaped proxy, not an LF/Rosenberg model with named open quotient, speed quotient, and return phase.
openingis used as waveform skew in generated glottis records. That field name is too generic for what it does here.- Source injection bypasses boundary impedance; it is simply added into node outgoing waves.
Verdict: acceptable as a bounded placeholder. Rename or split glottal shape fields, then implement a real glottal flow source with impedance coupling.
Turbulence And Release Source
What it simulates: pressure-driven constriction noise and a short opening transient.
Current state: partially coherent. Sustained turbulence now has a local pressure-drive term, and release pressure can observe node incident pressure. Graph turbulence lowering now emits named closure, release, reservoir, pressure-drive, gate, burst, and release-pulse locals instead of hiding the whole pressure-flow model inside one source expression.
Smells:
- The aperture gate constants
0.85,0.04, source gains0.62,1.2,0.58, and pressure-drive0.50 + 1.50 * pressureare not owned by a named physical model. They are now at least localized in graph source lowering. - Release still detects local control opening and stores pressure at the source owner, not as a true two-sided constriction/contact state inside the scatter graph.
- There is no explicit upstream/downstream reservoir split across a sealed constriction.
Verdict: keep as a named pressure-informed source. Do not call this plosive modeling yet. The next real primitive is a closure/contact owner that can store pressure on one side of a severe constriction and release it through the scattering graph.
Segment Loss
What it simulates: propagation loss plus extra loss in narrow/contact-like sections.
Current state: directionally correct and audibly important. It stopped closed sections from preserving wave energy as immortal buzz.
Smells:
- The
0.02area constant is an exposed scar. It says "area-dependent loss" but not whose units or morphology it belongs to. - Loss is amplitude-only. There is no frequency-dependent wall/radiation loss model.
Verdict: keep the authority, parameterize it. Path or material records should own contact loss scale and frequency-dependent wall loss.
Radiation
What it simulates: boundary flow emitted from lip, nostril, beak, or other openings.
Current state: less wrong after cutting generated lip double gates and moving the graph output constants into one named radiation impedance expression. Radiation now reads boundary flow, generated tract lip opening is path geometry, and emitted graph Faust names per-port reference area, differentiation, high-pass cutoff, admittance, and flow.
Smells:
- The graph no longer has inline
30/70raw/high-pass blend orarea + 1.0admittance denominator, but the helper still uses compact empirical kind constants. Those are now in one organ, not scattered through graph lowering. RadiationBoundaryReflectionExpressionis useful for standalone radiation openings, but dangerous when the path end area already represents aperture.- Radiation still has no frequency-dependent load matched to tube radius and characteristic impedance.
Verdict: keep boundary-flow readout and the named primitive. Replace the empirical primitive with a radius/characteristic-impedance model once source and closure storage stop dominating the audible failure.
Generated PT-Style Tract
What it simulates: a compact human vocal tract morphology expressed through neutral graph records.
Current state: much healthier. It no longer relies on PT-shaped runtime modules, and the accepted utterance smoke now passes without cross-word collapse.
Smells:
- It still generates many area terminals and many source terminals. That is a compiled approximation of continuous morphology, not a beautiful abstraction.
- Nasal and lip reflection both bind to
lip_reflection, matching PT pressure but not a general anatomy model. tract_motionis not applied to graph area controls.
Verdict: good pressure fixture, not final architecture. Continuous/interpolated source and area sampling would make it less PT-grid-shaped.
Tests And Metrics
What they simulate: falsification pressure, not hearing.
Current state: better. Cross-utterance separability caught the metric lie. Static closure-release no longer pretends a sealed static constriction has time-history release energy.
Smells:
- Smoke floors are still survival rails, not acceptance criteria.
- Artifact WAVs are generated, but the harness has no durable listening verdict.
- Speech-band/articulation metrics are still weak proxies.
Verdict: keep the new separability check. Add windowed phone/gesture reports as first-class artifacts because they exposed the nasal and rounded-vowel failures faster than global scores.
Cut List
Cut or replace these before piling on more tuning:
- Unowned segment contact-loss constant
0.02. - Turbulence/release scalar constants without a named pressure-flow model.
- Nearest-node source attachment as the only interior source placement model.
- Any new vocal feature added before the primitive-collapse and flow-probe
plan in
docs/acoustic-graph-proprioception.md.
Cut in the 2026-05-28 N-port pass:
- Generic connection fallback pressure/admittance blend.
- Shape-specific three-port branch scattering path.
Cut in the 2026-05-28 radiation pass:
- Inline graph radiation
30/70raw/high-pass blend. - Inline graph radiation
area + 1.0admittance denominator.
Keep List
These foundations deserve to stay:
- Path graph with typed terminals.
- KL/PT traveling-wave convention for two-port and nasal branch scatter.
- Source terminals not owning topology.
- Velum as side-port area, not global branch damping.
- Generated lip opening as path-end geometry, not output double gate.
- Area-dependent segment loss as a named owner, once parameterized.
- Cross-utterance separability as a hard guardrail.
Next Coherent Build
The next implementation should not be another gain pass. The general passive N-port junction now exists; the remaining build targets are closure storage and radiation impedance:
- Closure reservoir: a stateful pressure store attached to a constriction owner, charged while local area is sealed and released through the graph when area opens.
- Radiation impedance: upgrade the named empirical local model to a radius/characteristic-impedance model.
Until those exist, the graph can imitate some Pink Trombone pressure, but it is not yet a clean general vocal-acoustic machine.
2026-05-28 N-Port Connection Cut
The first audit cut replaced connection fallback lowering with one general area-scattering law for all non-bypass acoustic connections. This removes two old authorities at once: the pressure-like fallback expression and the PT-shaped three-port branch recognizer.
Latest utterance artifact:
artifacts/parity/pink-trombone-utterance-logmel/20260528T113334390.
The change keeps smoke tests passing. thrombosis improves slightly to cosine
0.4555, while mama/papa remain differentiated (candidateLogMel=0.2469
versus reference 0.3791) but still under-articulated and quiet.
Latest static artifact:
artifacts/parity/pink-trombone-logmel/20260528T113334415.
Static fixtures still pass; closure-release remains sealed-static evidence
with peak 0.000009.
This is not a victory lap. It is a floor repair. The next real smells are still radiation impedance constants, graph-native closure reservoir state, and source impedance at junction nodes.
2026-05-28 Radiation Primitive Cut
The second audit cut moved graph radiation coloring and admittance into a named local primitive. The emitter now names per-radiation-terminal:
graph_radiation_reference_area_*graph_radiation_differentiation_*graph_radiation_highpass_*graph_radiation_admittance_*graph_radiation_flow_*
This removes the inline flow * 0.30 + highpass(flow) * 0.70 expression and
the inline area + 1.0 admittance denominator from graph lowering. The current
primitive is still empirical, but it has one owner.
Latest utterance artifact:
artifacts/parity/pink-trombone-utterance-logmel/20260528T113949841.
mama/papa/thrombosis cosine is 0.7648/0.8550/0.4556. mama/papa
candidate separability is 0.2044 versus reference 0.3791, so this cleanup
does not fix articulation.
Latest static artifact:
artifacts/parity/pink-trombone-logmel/20260528T113949813.
Static fixtures still pass; closure-release remains sealed-static evidence
with peak 0.000011.
The conclusion is plain: radiation needed an owner, but the voice still needs graph-native closure pressure storage and source impedance before it will stop leaning on boundary color.
2026-05-28 Graph Source Reservoir Naming
The third audit cut did not change the intended source math. It moved graph turbulence/release lowering out of one dense expression and into emitted per-source locals:
_closure_release_reservoir_pressure_drive_release_pressure_noise_gate_release_gate_sustained_burst_noise_release_pulse
Latest utterance artifact:
artifacts/parity/pink-trombone-utterance-logmel/20260528T114618213.
Latest static artifact:
artifacts/parity/pink-trombone-logmel/20260528T114618202.
Metrics match the prior radiation-primitive pass, which is the expected result:
this was an ownership/observability cut, not a gain change.
The next change can now target the reservoir/source-impedance law directly. The remaining bad smell is not that the source code is unreadable; it is that the model still has only a source-local reservoir instead of a contact-aware upstream/downstream pressure state.
2026-05-28 Reservoir Charging Sign
The source-local closure reservoir now charges from positive local pressure only:
max(0, incident_pressure) * closure
The previous abs(incident_pressure) * closure let alternating buzzer pressure
charge a plosive reservoir equally in both phases. That is not pressure storage;
it is rectified vibration. The new law is more physical, but current parity
metrics did not move:
- utterance artifact
artifacts/parity/pink-trombone-utterance-logmel/20260528T114903720; - static artifact
artifacts/parity/pink-trombone-logmel/20260528T114953503.
Keep the sign fix. The non-result strengthens the diagnosis: the real missing primitive is still a contact-aware upstream/downstream closure state, not a source-local scalar reservoir.
2026-05-28 Directed Reservoir Drive
For turbulence/release sources attached to ordinary two-port path nodes, the reservoir now charges from directed upstream-minus-downstream pressure:
max(0, upstream_incoming - downstream_incoming)
Boundary or non-two-port source nodes fall back to positive local incident pressure. This lets a source-local plosive at least read the graph's traveling wave direction instead of only a node-average pressure.
Latest utterance artifact:
artifacts/parity/pink-trombone-utterance-logmel/20260528T115452268.
Latest static artifact:
artifacts/parity/pink-trombone-logmel/20260528T115719882.
Metrics still do not move. That is useful and damning: the source-local reservoir sign/direction is no longer the main missing piece. The graph needs a contact/closure primitive that participates in scattering and flow, not a separate source pulse hoping the node will turn it into articulation.
2026-05-29 Radiation Boundary Load And Node-Law Cut
Radiation terminals are no longer output-only observers at one-port boundaries.
For a boundary radiation node, graph lowering now emits a named
graph_radiation_boundary_load_* value from the radiation admittance,
differentiation, and loss. The reflected boundary wave is filtered before it
returns to the delay line:
reflected = raw * (1 - load) + lowpass(raw, radiation_cutoff) * load
The output still reads boundary flow as incoming - outgoing, but the outgoing
wave now carries radiation load back into the tract. This is not a full
frequency-dependent radiation impedance model. It is the first graph-owned
load term that changes the pressure field instead of only coloring the output.
The remaining unconnected multi-port node fallback was also cut. Multi-port
nodes not owned by an explicit connection now use the same area-scattering
wave convention and emit node area-energy probes. The old
graph_node_pressure_* expression should not reappear as a quiet fallback.
Latest graph probe:
artifacts/parity/pink-trombone-graph-thrombosis-probes/20260529T112805836.
The new lip boundary load peaks at 0.516196, lip flow remains alive at
0.586863, modal source out is 1.234403, and the sibilant area-34 source
out is 0.099318.
Latest utterance artifact:
artifacts/parity/pink-trombone-utterance-logmel/20260529T112806745.
The accepted-utterance diagnostic still fails articulation. thrombosis
cosine is 0.3287, RMS ratio is 0.0770, active ratio is 0.3151, and
silence mismatch is 0.7205. This confirms the boundary-load cut is a
coherence repair, not the missing speech primitive.
Next pressure remains contact as constrained-flow history and source impedance as a port law. Do not interpret this pass as permission to restore global drive or output EQ.