GameCult/AquaSynth
Watch
9
0
Fork
You've already forked AquaSynth
 0
Code Issues Pull requests 4 Projects Releases Packages Wiki Activity Actions

Bound PT utterance render diagnostics

Browse source
This commit is contained in:
Meta Vens 2026-05-29 01:57:38 +01:00
commit d8a33f7e6b
4 changed files with 61 additions and 10 deletions
Download patch file Download diff file Expand all files Collapse all files

16
src/AquaSynth.Faust/FaustCompiler.cs
View file

@ -284,9 +284,19 @@ public static class FaustCompiler
foreach (var argument in arguments) start.ArgumentList.Add(argument);
using var process = Process.Start(start) ?? throw new InvalidOperationException($"failed to start `{fileName}`");
var stdout = process.StandardOutput.ReadToEndAsync(cancellationToken);
var stderr = process.StandardError.ReadToEndAsync(cancellationToken);
await process.WaitForExitAsync(cancellationToken);
var stdout = process.StandardOutput.ReadToEndAsync();
var stderr = process.StandardError.ReadToEndAsync();
try
{
await process.WaitForExitAsync(cancellationToken);
}
catch (OperationCanceledException) when (!process.HasExited)
{
process.Kill(entireProcessTree: true);
await process.WaitForExitAsync(CancellationToken.None);
throw;
}
return (process.ExitCode, await stdout, await stderr);
}
}

1
state/evidence.jsonl
View file

@ -215,3 +215,4 @@
{"evidence":"A visual review of the Warbling White-eye comparison showed the previous best candidate still failed the reference: its contours were piecewise straight and its harmonic ladder was too strong. Added `interp=smooth` control curves lowered to Faust `seg_smooth` smoothstep segments, then rerendered the Warbling White-eye audition with smooth pressure/opening/tension/stiffness gestures and a low-drive single-labium candidate. `artifacts/listening/warbling-white-eye-20260528T220500961/06-smooth-low-drive.compare.png` now shows curved motif contours and much weaker harmonics, though it is still only a rough birdlike gesture, not a species-quality match.","claim":"Smooth realtime gesture semantics are required for vocal modeling. Straight spectral contours are often authored linear control curves leaking through the physical model, so future bird/human fitting should use smooth motor gestures before adding morphology changes or spectral compensators.","timestamp":"2026-05-28T22:50:00+01:00","topic":"smooth-vocal-gestures"}
{"evidence":"Created `research/vocal-source-grounding` with downloaded/captured references on MEAD source physics, syrinx modeling, two-mass vocal fold modeling, self-oscillating vocal fold experiments, Birkholz 2024 pressure losses, and Faust physical modeling primitives. The grounding pass found `drive` is currently split-brained: it scales reservoir pressure, phonation gate, and modal/displacement brightness. Literature points instead to pressure as energy input, tissue mass/stiffness/damping/collision as source morphology, aperture/area-dependent flow loss as pressure-flow shaping, and tract/radiation load as spectral/acoustic coupling.","claim":"The next valve-law cut should demote generic `drive` into a calibration/compatibility alias, then make harmonic richness and ease of oscillation emerge from explicit morphology, pressure-flow loss, collision/contact, damping, and graph load rather than a timbre rescue knob.","timestamp":"2026-05-28T23:12:00+01:00","topic":"vocal-source-grounding"}
{"evidence":"Demoted graph tissue-valve `drive` from split-brained vocal authority into a parser compatibility alias for `flow_scale`, added explicit `flow_scale`, `tissue_loss`, `aperture_shape`, and `flow_loss` source-port fields, and changed Faust graph lowering so reservoir pressure is pressure-owned while oscillation threshold/Q, velocity decay, aperture law, and pressure-flow resistance come from physical morphology/load controls. Canonical human, bird, and alien graph patches now use the new fields. Focused parser/export/render tests passed, class-sliced non-PT suites passed, and fresh Warbling White-eye audition artifacts were written under `artifacts/listening/warbling-white-eye-20260529T002500-drive-demotion`. The PT accepted-utterance diagnostic renderer is unstable in this run: the mama artifact writes, then the testhost exits with code -1 before reporting a managed assertion.","claim":"Graph vocal source energy, threshold, and harmonic behavior now have separate inspectable owners. Future bird/human tuning should move through pressure, stiffness/tension, damping/tissue loss, aperture/flow resistance, collision, radiation, and graph load instead of reviving a generic vocal drive knob. The next cleanup should inspect the remaining low-frequency flow floor/radiation normalization and the PT diagnostic testhost crash, not use the crash as permission to restore a generic drive knob.","timestamp":"2026-05-29T00:20:00+01:00","topic":"graph-tissue-valve-drive-demotion"}
{"evidence":"Rechecked the PT accepted-utterance diagnostic after clearing stale child `dotnet` processes. The focused test passes with enough wall-clock budget; fresh artifact `artifacts/parity/pink-trombone-utterance-logmel/20260529T005420700` records render timings of about 31s for mama, 45s for papa, and 59s for thrombosis. `PinkTromboneAcceptedUtterancesReportGraphLogMelParityWhenFaustIsInstalled` now writes generated DSP/control artifacts before rendering, applies a three-minute per-utterance cancellation budget, and writes render stdout/stderr/timing. `FaustCompiler.RunAsync` now kills the entire child process tree when cancellation fires. Focused PT utterance plus graph Faust validation/render tests passed: 3 tests in 2m17s.","claim":"The previous PT testhost failure is not currently a reproducible managed graph-source crash; the real process smell was unbounded external Faust/dotnet render work with poor pre-crash artifacts. Current acoustic evidence points to underpowered/weakly articulated radiated flow, not a loud low-frequency floor, so next tuning should use graph probes rather than restore the generic drive authority.","timestamp":"2026-05-29T01:00:00+01:00","topic":"pt-utterance-process-observability"}

24
state/scratch.md
View file

@ -2,6 +2,30 @@
Current slice:
- PT utterance diagnostic crash recheck: the focused test is reproducible as
slow but passing when stale child `dotnet` processes are cleared and the test
is run with enough wall-clock budget. Fresh verified artifact:
`artifacts/parity/pink-trombone-utterance-logmel/20260529T005420700`.
Per-utterance render timing now writes `render.txt` beside each candidate
before/after the Faust render path: `mama` ~31s, `papa` ~45s, `thrombosis`
~59s. The previous two-minute attached shell timeout was not a valid crash
witness.
- Harness/process fix: `PinkTromboneAcceptedUtterancesReportGraphLogMelParityWhenFaustIsInstalled`
now writes `candidate-graph.dsp` and `controls.csv` before rendering, wraps
each external Faust render in a three-minute per-utterance timeout, and writes
stdout/stderr/timing to `render.txt`. `FaustCompiler.RunAsync` now kills the
whole child process tree when cancellation fires, so timed-out `faust` or
nested `dotnet run` work cannot keep poisoning later runs.
- Current acoustic evidence after the source-law cleanup is weakness, not a
loud low-frequency floor: latest utterance RMS ratios are `mama` 0.0876,
`papa` 0.1509, `thrombosis` 0.0673. `thrombosis` still fails the diagnostic
cosine floor with cosine 0.3318. Next vocal work should inspect source-flow
strength/articulation and radiation transfer with debug probes, not restore a
generic `drive` knob.
- Verification:
`dotnet test tests\AquaSynth.Dsl.Tests\AquaSynth.Dsl.Tests.csproj --no-restore --disable-build-servers -p:UseSharedCompilation=false -p:BuildInParallel=false --filter "PinkTromboneAcceptedUtterancesReportGraphLogMelParityWhenFaustIsInstalled|FaustCompilerRendersTissueValveGraphWhenInstalled|FaustCompilerValidatesAcousticPathGraphWhenInstalled" -v minimal`
passed: 3 tests in 2m17s.
- Research-guided machine cut: graph segment delay no longer clamps physical
segment delay to a full sample. `WaveClockPolicy` now owns the minimum legal
delay: unit grid stays at 1 sample, half-sample grid and first-order Thiran

30
tests/AquaSynth.Dsl.Tests/PinkTromboneLogMelParityTests.cs
View file

@ -1,4 +1,5 @@
using System.Globalization;
using System.Diagnostics;
using System.Text;
namespace AquaSynth.Dsl.Tests;
@ -111,9 +112,27 @@ public sealed class PinkTromboneLogMelParityTests
{
var reference = referenceRenderer.RenderUtterance(fixture.Id, fixture.ControlPoints, fixture.DurationSeconds);
var source = AutomatedGraphSource(fixture);
var candidate = await FaustCompiler.RenderAsync(
source,
new FaustRenderOptions(reference.SampleRate, reference.Samples.Length / (float)reference.SampleRate));
var fixtureDir = Path.Combine(artifactDir, fixture.Id);
Directory.CreateDirectory(fixtureDir);
File.WriteAllText(Path.Combine(fixtureDir, "candidate-graph.dsp"), source);
File.WriteAllText(Path.Combine(fixtureDir, "controls.csv"), UtteranceControlCsv(fixture));
var stopwatch = Stopwatch.StartNew();
using var renderTimeout = new CancellationTokenSource(TimeSpan.FromMinutes(3));
FaustRender? candidate;
try
{
candidate = await FaustCompiler.RenderAsync(
source,
new FaustRenderOptions(reference.SampleRate, reference.Samples.Length / (float)reference.SampleRate),
cancellationToken: renderTimeout.Token);
}
catch (OperationCanceledException)
{
Assert.Fail($"Faust render timed out for `{fixture.Id}` after {stopwatch.Elapsed.TotalSeconds:0.0}s; artifacts: {fixtureDir}");
return;
}
stopwatch.Stop();
Assert.NotNull(candidate);
Assert.True(candidate.Samples.Length > 0, $"{candidate.Stderr}{Environment.NewLine}artifacts: {artifactDir}");
@ -124,13 +143,10 @@ public sealed class PinkTromboneLogMelParityTests
var report = UtteranceReport(fixture, comparison);
reports.Add(report);
var fixtureDir = Path.Combine(artifactDir, fixture.Id);
Directory.CreateDirectory(fixtureDir);
WriteWav(Path.Combine(fixtureDir, "reference-pink-trombone.wav"), reference.Samples, reference.SampleRate);
WriteWav(Path.Combine(fixtureDir, "candidate-graph.wav"), candidate.Samples, candidate.SampleRate);
File.WriteAllText(Path.Combine(fixtureDir, "candidate-graph.dsp"), source);
File.WriteAllText(Path.Combine(fixtureDir, "report.txt"), report);
File.WriteAllText(Path.Combine(fixtureDir, "controls.csv"), UtteranceControlCsv(fixture));
File.WriteAllText(Path.Combine(fixtureDir, "render.txt"), $"seconds={stopwatch.Elapsed.TotalSeconds:0.000}{Environment.NewLine}stdout:{Environment.NewLine}{candidate.Stdout}{Environment.NewLine}stderr:{Environment.NewLine}{candidate.Stderr}");
if (comparison.LogMelCosineSimilarity < UtteranceGraphSmokeCosineFloors.GetValueOrDefault(fixture.Id, 0.1f))
{