Skip to the answer
Audio MIDICONVERTER / FIELD NOTESOpen converter

INDEPENDENT BOUNDARY SWITCHES / OCTOBER 18, 2026

MP3 to MIDI.
One boundary. Then the other.

With the same authored 440 Hz source clock and global envelope, the actual WAV / MP3 MIDI counts were: neither reset 1 / 1, first boundary only 2 / 2, second only 2 / 2, both 3 / 3. The single-boundary cells are new offline CPU inference; the endpoints retain exact-input captures. These conditional counts do not identify intended musical attacks.

Fixed source clock / Two reset switches / Actual MIDI fields

The middle cells
were not measured before.

B / FIRST BOUNDARY ONLY / NEW INFERENCE2 / 2 events.WAV / MP3. No second restart, no per-slot fade or added silent gap.
C / SECOND BOUNDARY ONLY / NEW INFERENCE2 / 2 events.WAV / MP3. Phase remains continuous through the first boundary.
Actual re-parsed MIDI counts. The selected source, slot length, amplitude, global envelope, model and extraction settings are fixed across this matrix.
CellFirst resetSecond resetWAV / MP3 eventsProvenance
A / Neither boundary resetsnono1 / 1Retained Oct17 anchor
B / First boundary onlyyesno2 / 2New Oct18 offline CPU inference
C / Second boundary onlynoyes2 / 2New Oct18 offline CPU inference
D / Both boundaries resetyesyes3 / 3Retained Oct17 anchor

A / Neither boundary resets

First reset: no. Second reset: no. WAV / MP3 events: 1 / 1. Retained Oct17 anchor.

B / First boundary only

First reset: yes. Second reset: no. WAV / MP3 events: 2 / 2. New Oct18 offline CPU inference.

C / Second boundary only

First reset: no. Second reset: yes. WAV / MP3 events: 2 / 2. New Oct18 offline CPU inference.

D / Both boundaries reset

First reset: yes. Second reset: yes. WAV / MP3 events: 3 / 3. Retained Oct17 anchor.

The preceding exact-cycle and off-cycle comparison changed both internal reset boundaries together. It established a source-byte identity in one construction and a different output in another, but could not assign a conditional result to either boundary alone. Here B and C supply those missing source interventions. A and D remain labelled endpoints from the preceding off-cycle case, not newly discovered results.

There are 4 fresh input inferences, 4 retained prior inputs and 0 same-input reuses within this run. The paired WAV and MP3 encodings are not independent statistical replicates. Each reported count comes from its corresponding saved MIDI, and full pitch, timing and velocity fields are retained even when totals happen to match.

First-only versus second-only / WAV: counts 2 and 2. Complete parsed note fields identical: no. Actual MIDI bytes identical: no. These are separate checks, not conclusions drawn from matching totals.

First-only versus second-only / MP3: counts 2 and 2. Complete parsed note fields identical: no. Actual MIDI bytes identical: no. These are separate checks, not conclusions drawn from matching totals.

Actual exported MIDI counts for neither, first-only, second-only and both phase reset controls, with retained anchors separated from new inference
Dark bars mark new inference. Muted bars mark retained exact-input anchors. Open the count chart.

Hold the other switch.
Read the conditional difference.

Descriptive differences in actual exported event counts. Input counts A through D are separately recorded for WAV and MP3 in results.json.
ComparisonFormulaWAVMP3
First reset, second offB - A11
First reset, second onD - C11
Second reset, first offC - A11
Second reset, first onD - B11
Cross-difference(D - C) - (B - A)00

First reset, second off

Formula: B - A. WAV / MP3: 1 / 1.

First reset, second on

Formula: D - C. WAV / MP3: 1 / 1.

Second reset, first off

Formula: C - A. WAV / MP3: 1 / 1.

Second reset, first on

Formula: D - B. WAV / MP3: 1 / 1.

Cross-difference

Formula: (D - C) - (B - A). WAV / MP3: 0 / 0.

Enabling the first reset with the second disabled gives B - A. With the second enabled, the corresponding contrast is D - C. Enabling the second reset with the first disabled gives C - A; with the first enabled it gives D - B. The result JSON names every input MIDI filename and its actual parsed count. No curve or expected number of splits is substituted for an export.

The cross-difference (D - C) - (B - A) is 0 for WAV and 0 for MP3. It is arithmetic on this fixed construction, not a statistical interaction test, a probability, an accuracy score or evidence of a universal phase rule. Equal conditional count changes would still not prove equal start times, velocities or musical meaning.

The switches are separate generator decisions, but their downstream phase states need not be independent. With only the first restart, the oscillator retains that phase origin through the final slot. With both restarts, its origin changes again at the second boundary. Turning on the first switch can therefore change the phase state arriving at the second boundary even when the second switch is held fixed. This matrix separates the decisions; it does not claim to inject two identical, independent acoustic impulses.

A source boundary
is not a note label.

The first internal boundary is sample 22051, or 22051 / 22050 = 1.000045351473923 seconds. The second is sample 39692, or 39692 / 22050 = 1.800090702947846 seconds. These are selected construction positions. They are not onset times measured from a musical performance.

Actual WAV MIDI event intervals plotted against the two fixed source boundary positions, with enabled resets distinguished
Vertical lines are source positions, not assigned musical attacks. Bars use actual MIDI start and duration fields. Open the interval chart.

For each parsed event we retain two signed distances: startSeconds - firstBoundarySample / sampleRate and startSeconds - secondBoundarySample / sampleRate. The actual MIDI start field and exact source sample position supply the inputs. Negative means the exported event starts before that source boundary; positive means after. This is not a nearest-attack assignment, and we set no tolerance that turns a close distance into a correct note.

The display below rounds seconds to six decimal places. The result JSON preserves full numbers, ticks, duration ticks, PPQ and velocity; equality checks use full fields and byte hashes, not the rounded labels. An increase in the total does not by itself identify which interval was split, whether one vanished, or what a performer intended.

Keep the clock.
Change only the enabled resets.

The authored oscillator uses selected frequency 440 Hz at 22050 samples per second, with 70560 total samples. Active audio begins at sample 4410. Three touching slots each contain 17641 samples. One global envelope spans the full active interval, with 220-sample edges and nominal amplitude multiplier 0.35. No internal fade, positive silent gap or RMS normalization is introduced. Equal nominal amplitude is not a claim of equal RMS; each actual RMS is recorded.

The source evaluates sin(2π × (((i - origin) × 440) mod 22050) / 22050), multiplied by the global edge envelope and quantized to PCM16. The phase origin starts at the first active sample. At an enabled internal boundary it becomes that boundary's sample index; at a disabled boundary it stays at its most recent origin. The final active endpoint and the envelope remain fixed across all four cases. All construction slots are bookkeeping intervals, not asserted audible attacks.

For every conditional source pair, the independent checker verifies that PCM before the toggled boundary is unchanged and independently regenerates every sample and the WAV header. The measured changed-sample counts are neither to first-only: 35281; second-only to both: 17641; neither to second-only: 17640; first-only to both: 17639. Those values compare actual quantized source arrays. They are not human hearing thresholds, recommended edit settings or a count of correct notes.

The prior anchors are accepted only when regenerated input bytes, decoded mono float32 samples and model hashes match their Oct17 records. Their original capture timestamps remain intact. The new middle cells run the unchanged installed model on the offline CPU backend; the public converter and all older experiment bodies remain untouched.

Keep the capture
and its date together.

Basic Pitch 1.0.1, TensorFlow.js 3.21.0, MIDI writer 2.0.28, v24.19.0 on win32; CPU backend. Native FFmpeg encodes MP3 at 128 kbps and decodes each file to mono float32 at the selected sample rate. The onset and frame thresholds stay at 0.25 and 0.25; extraction minimum stays at 5 frames. All raw frame, onset and contour matrices are retained.

Model manifest SHA-256: 1ed1aaee3409ec1dc098c8b01f430c0911f6fe9412e7af8086750f9e8f302f68. Combined model weights: b142a95737a52e1e412d5f92e73d8bb80dfe8d04941acc0702f11f4524fb377c. Unchanged converter source: e42f9b0b60bb6fd73b2539a1ae9ca3b09f34ab874e35423162b384df68c19660. A separate replay re-extracts the capture, rebuilds each MIDI and requires its actual file hash and parsed fields to match the saved export. A retained row does not become a fresh experiment merely because this page is new.

MP3MIDITool's repair guidance distinguishes fragments belonging to a held note. That observed distinction motivates this evidence sequence, not a claim that its converter was tested here. The fixed comparison pages were read for workflow and gaps; no site-wide absence or current ranking claim follows from that read.

This is synthetic-source, offline CPU evidence. It is not a real-instrument recording, listening test, browser inference, physical-phone conversion, DAW result, musical-intent classifier or accuracy percentage. No source recording belonging to a visitor was used or uploaded. The previous joining experiment and repeated-note counterexample retain their separate rules and reciprocal links.

Read the event fields.
Not just the headline count.

Every line below is re-parsed from its actual Standard MIDI export. Display rounding is for reading only. The source reset choices, complete model capture and export remain separately downloadable, so a reader can inspect source geometry without treating a note total as the whole result.

Download the controls.
Recheck the inference.

Full results and provenance contain conditional formulas, source boundary positions, original timestamps, hashes and event fields. Reproduction code runs in this project's locked environment as node scripts/measure-boundary-isolation.cjs. Prior Oct17 evidence is required for anchor validation. The --replay option refuses missing captures instead of manufacturing inference.

The page reports actual measurements rather than predicted results. Its conditional changes are limited to the declared source, sample clock, phase arithmetic, envelope and model path. They do not supply a safe musical rule for other audio. The preceding phase-boundary page remains available with its original exact-cycle and off-cycle claims.