Reed Study 005 — preferred voices and intentional notes

This build gives Charlie's two preferred comparisons a playable keyboard, repeatable musical gestures, and shorter listening comparisons. Opening is the opening-only branch from the matched-pressure experiment. Loose is the full loose-lip branch. Contour is an optional intermediate resonance/damping candidate. Charlie has not yet evaluated these new performances.

Open play.html for the instrument and focus.html for focused listening. The earlier instrument remains at index.html and the earlier laboratory at lab.html. No external messages are sent; feedback and recordings download locally.

What to try

1. Start with Opening and Loose at the reference Breath 80%. Play a few separated notes, then connect notes across the octave. Does the preferred character survive in a melody, and which voice belongs in the foreground or underneath something else?
2. Try the two Beginning choices. Breath-led is the default. Tongued adds brief contact while the same pressure curve rises. Listen for the arrival you want; a more elaborate physical action is not automatically more musical.
3. Try Short and Breath endings. Short falls promptly; Breath exposes more of the physical tail and its pitch movement. The focused page's Release button makes this comparison brief and repeatable.
4. Try Push and recover. The phrase returns breath to its reference value. Does it recover in a way you can use? If it gets stuck or sounds unexpectedly sharp, save the notes/settings file; it contains the exact voice and conditions.
5. Contour is a candidate for more resonance with less damping. It is not a replacement for either preferred voice.

Voice definitions and breath

Opening: rest opening 0.4186 mm; pressure scale Pc 6756.75 Pa; reed resonance 1031.988452 Hz; damping ratio 0.580.
Loose: rest opening 0.4186 mm; Pc 4478.25 Pa; reed resonance 1031.988452 Hz; damping ratio 0.468.
Contour: Loose's opening and pressure scale; reed resonance 1699.872838 Hz; damping ratio 0.286.

Opening preserves the lip-85% reference mechanics with only the rest opening changed to its lip-15% value. Loose preserves the full lip-15% mapping. Both use the earlier Supple settings elsewhere. Contour changes resonance and damping together; it is a musical candidate, not an isolation experiment identifying which of those two changes causes a preference.

Breath now directly sets gamma = supply pressure / Pc. The reference is gamma 0.8, displayed as 80%, and the range is 0 to 2.1. This is a relative pressure scale, not loudness or percent of maximum. At reference, Opening receives 5405.4 Pa and Loose/Contour 3582.6 Pa. Maximum supply is 14189.175 Pa for Opening and 9404.325 Pa for Loose/Contour. The former 9000 Pa slider ceiling was a control limit; the expanded range is explicit and tested for finite operation. It is not a physiological claim.

Pc is the unconstrained reed's closing-pressure scale. Because the retained mouthpiece contact law resists closure, gamma 1 does not imply guaranteed sealing. The legacy contact/closure issue raised in the reviews has not been silently relabeled as fixed.

Pitch and physical state

Pitch still comes from the simple cylindrical air-column delay. There is no substitute pitch oscillator or output pitch shifter. Each voice has its own measured G3–G4 tables for Standard and Higher rate at 44.1, 48, and 96 kHz output. Reference generation uses a fresh 60 ms linear pressure rise to gamma 0.8, followed by a held measurement. The actual performance controller uses a smooth pressure rise and is tested independently.

The tables tune around the reference conditions; they do not follow pitch during a breath push. Connected notes preserve reed and tube state. A note arriving during a release also inherits residual state, so rapid repetitions can differ from a fresh note. Voice/rate changes finish the current release and instantiate the next voice at rest.

Starts and endings

A fresh note begins from rest, with zero samples before its declared noteOn. No preparatory note is played or simulated as a hidden warmup. Breath rises over 60 ms; the optional tongue pulse acts during that rise. An 8 ms output ramp also softens reentry into a partly faded release.

In the G3/C4/G4 attack probes, a 10 ms window first exceeded 10% of held RMS about 90–110 ms after noteOn. That is the model's acoustic buildup, before any browser/device latency. Whether it feels suitably responsive needs Charlie's playing feedback.

The selected Short policy reduces pressure over 70 ms, with no closing tongue pulse, and applies an explicit 60 ms output fade. The physical engine continues for a 500 ms tail before resetting at rest. The Breath policy falls over 340 ms, has no additional release output fade, and permits an 850 ms physical tail. Tests inspect the engine output before the controller fade as well as the heard output.

During development, a release tongue pulse produced a sharp C4 tail. Removing it reduced that problem; the shorter output fade limits the later falling-pitch portion. This is a deliberate performance policy. It does not prove the reed mechanics alone now create an ideal release. The audible Short-versus-Breath pair changes both the pressure envelope and output envelope; it is not an isolated fade-only experiment.

Listening comparisons and reproducibility

Ten fixed-gain mono WAVs cover five questions: preferred voices, beginnings, endings, Contour, and push/recovery. Each pair uses the same event score. Attack / Held / Release select explicit time intervals; bounded repeat loops stay inside the selected interval. The player marks note and control events. Optional RMS matching measures the heard excerpt, only attenuates, and refuses very quiet or more than 18 dB-different pairs.

Playback applies a 5 ms edge fade to copied excerpts to avoid crop clicks. Original WAV downloads are unchanged. Raw excerpt downloads omit player matching, listening volume, and crop fades; they still contain the performance controller's original note envelope. These are distinct from the engine's pre-envelope raw samples measured by the render diagnostics.

Live AudioWorklet playback, recording, replay, and offline export use PlayPerformance and the same event-score timeline. The manifest records exact controls, rate, seed, voice, regions and source hashes. render-diagnostics.json includes the controller traces and pre-envelope release measurements. verification.json independently decodes the delivered WAVs and records acceptance checks. Feedback exports retain the selected pair, region, loop and playback policy.

Body and output scope

The reed/body feedback engine remains the frozen backward-Euler, legacy-contact model, with mouth coupling and small deterministic flow noise off in this playable branch. Higher rate is the default: 352.8 or 384 kHz internal depending on output rate. Standard is available if a device struggles. This preserves the observed preference for the backward-Euler branch while acknowledging its rate-dependent damping; it does not establish numerical convergence.

The Aftertouch handoff's separation of excitation, resonator, output observation and voicing is useful here. We did not transfer its beam geometry laws into the wind model. A bounded audit changed only the output high-pass from 950 to 600 Hz and checked the reed/air-column states for exact equality. The detailed result is in radiation-audit.json. The alternative raises held output by about 3.1 dB and lowers the measured spectral centroid in the tested notes, while physical feedback is unchanged. The playable default remains 950 Hz; this is an output-voicing option for later listening, not a body redesign or a physics repair.

Questions for Claude, Argon, and Grok

- Evaluate the note controller as a musical policy, separately from the contact law. Is there a small mechanically motivated way to improve the first and last 100 ms without losing the two preferred steady voices?
- With opening and pressure now explicit, which single next control would expose useful variation while preserving pitch and repeatability? Avoid rebundling several mechanical changes into a misleading physical label.
- Should the next mechanical experiment vary sealing/leak independently of contact onset, using these exact voices and scores? The old leaky branch remains a documented reference.
- Is an isolated output-voicing comparison useful before expanding the acoustic body? Any proposed output change should preserve and verify the physical feedback path.
- Please distinguish observed model behavior, hypotheses about acoustic realism, and listening preferences. The user should decide what is musically worth keeping.

Verification summary

An independent broad periodicity estimator checked 156 fresh default-start cases: all 13 notes, all three voices, both calculation rates, and 44.1/48 kHz output. Maximum measured error at reference breath was 2.355 cents. Separately, the frozen calibration generator covered 234 note/rate/voice cases, including 96 kHz output, with maximum reference error 0.340 cents using its calibration estimator. These are different measurement procedures, not conflicting claims of precision.

Connected-note, bend, pressure-push/recovery, finite-tail, exact pre-note silence, and recorded-replay checks passed. The tested live Worklet and offline score outputs match sample-for-sample at 44.1/48 kHz. Delivered WAVs were independently decoded and checked against declared durations, regions, voice settings, source hashes, and level measurements. No numerical fault resets, protection hits, or pressure-solver failures were reported for the curated performances. Interface tests use mocks to exercise held-key priority, asynchronous cancellation, recording, export identity, and lifecycle stops; they are not substitutes for device testing.

Limits

Numerical tests and DOM/Web Audio mocks cannot establish browser layout, autoplay, download behavior, or iPhone real-time performance. No real-browser/device QA was available in this environment. Node throughput measurements have scheduling outliers. These are synthesized model experiments; no acoustic-instrument fidelity or human listening approval is claimed. The free-reed/harmonica architecture remains separate future work.