# MP3 to MIDI: real recorded inputs, actual browser downloads, separate checks

Measurement date:2026-11-09, Asia/Seoul. This upgrades the existing homepage; it is not a new article URL, a new algorithm, a new polyphonic feature or a ranking/traffic forecast. The owner-provided search/DR/backlink/traffic figures are campaign inputs, not measurements from this experiment.

## Sources and changes

Three previously archived human acoustic-instrument microphone excerpts: GuitarSet1.1.0 solo/comping and Vienna4x22 piano. Both publisher records specify CC BY4.0. Exact sources, crop hashes and original full-file links are in source-provenance.json and ATTRIBUTION.md. The guitar source crops are2.74s and the stereo piano crop6s; no synthesizer, SoundFont or MIDI render replaces the recorded audio. Original archived WAVs, annotations, detector and prior seals remain unchanged.

One fixed FFmpeg/libmp3lame build encodes each original crop at192kbps,44100Hz, without normalization, then decodes to signed PCM16. The two guitar sources remain mono. The one piano MP3 is decoded once to stereo PCM and diagnostic mean/left/right WAVs are derived from that same decoded sample array. Mean is round((L+R)/2), with Python’s tie-to-even rounding; the channel diagnostics are offline controls, not extra page modes or proof of an acoustic phase/noise cause. encoding.json records actual encoder version/binary hash, sample counts and measured RMS values. Fixed encoder/build is required for the claimed byte hashes; other builds can change codec bytes and decoded samples.

Use unchanged production downsample/detectNotes/buildMidi functions from detector-snapshot.js. Guitar sources and decoded controls run settings1/2/3 as declared; piano controls run setting2. The reference contract stays exact integer pitch plus the nearest unused reference onset within150ms, inherited annotation order, one-to-one and no duration acceptance condition. Pitch-class/octave time-overlap candidates are reported separately. Maximum overlap ends a note before starting another at an identical timestamp. References include crop-crossing events. GuitarSet annotations are partly automated; the piano reference is hardware key-on/off with a fixed audio-minus-MIDI offset-0.63375s inherited from the archived onset calibration, not refitted to improve these predictions. Its validation residual reaches28.75ms, and key releases exclude pedal-held/acoustic decay. Do not treat these references as an infallible fresh listening transcription.

## Actual browser versus harness

All three encoded MP3s were selected with the actual in-app browser file chooser on the production homepage, converted at Balanced2, visibly completed, and downloaded using the actual .mid link. Those three unedited browser MIDI files are retained as browser-*.mid. Each is byte-identical to the corresponding FFmpeg/control harness MIDI in this particular run. browser-conversions.json records the operation and recovery of a file-selection timeout; a timeout is not an accepted conversion. Browser decoder sample rate was not independently inspected, and no general cross-browser equality, privacy network audit, listening study, DAW import or competitor inference is claimed.

## Reproduce the numerical results

Download the three *-source.wav files, reference.json, detector-snapshot.js, midi_reader.py, generate.py, convert.mjs, analyze.py and compare.py into one new directory, plus the three original browser-*.mid downloads if you wish to verify the browser/harness comparison. Install Node22+ and Python3 standard library; put your FFmpeg executable on PATH. The actual recorded encoder version and executable SHA256 are in encoding.json; byte reproduction here used the same binary/build, not a promise about every FFmpeg version.

```text
python generate.py --ffmpeg ffmpeg
node convert.mjs
python analyze.py
python compare.py
```

generate.py refuses to overwrite encoded/decode files, so use a clean directory. compare.py reads the three retained browser downloads; it does not simulate browser clicks or generate pretend browser outputs. If comparing a fresh manual browser test, retain it separately rather than replacing the originals, then parse its MIDI bytes independently.

The independent Python reader decodes the actual SMF bytes and checks the detector’s exported note data within3ms of tick rounding. Note ledgers expose reference and output pitch/start/duration. RMS-window counts separately recompute the same block downsampling,2048-sample window,512-sample hop and mode gate, without changing detection. Whole-crop RMS and per-window RMS gate counts are different metrics. comparisons.json separates timing, velocity, event count, onset-rule matches and file-byte identity. One fresh-directory run regenerated49 MP3/PCM/MIDI/CSV/JSON/SVG files byte-identically; the actual browser files are independently acquired and are not counted as regenerated by the offline script.

## Interpret without overclaiming

More events do not imply more correct attacks or simultaneous notes. A larger sensitivity value is a gate/grouping setting, not an instrument selector or polyphonic model. The real MP3 solo control has fewer exported events at More notes than Balanced; the chord control has a different response. PCM/codec and channel-mix changes can change timing or velocity even when a pitch sequence remains unchanged. These are observations on the named excerpts, not universal effects of MP3, recommendations for a best gain/channel, or a musical score for an entire instrument.

The piano left-channel onset match lies only7.708ms inside a150ms tolerance and its matching interval covers28.43% of the key-held reference duration; retained clock uncertainty limits that acceptance judgment. The guitar MP3 still has low octave candidates and an unsupported high event; the earlier lag unit/arithmetic audit remains sealed and is not reopened or called a production fix here. Different instruments also change performance, room, dynamics, pedal behavior and reference generation, so this is not instrument-only causal isolation. No commercial converter was run, ranked or found worse. The comparison range for editorial original observations is the actual ten reviewed Google bodies recorded in the research audit; it is not worldwide exclusivity.

Whole-page sentence auditing includes the first answer, all three evidence cards, methods, FAQ questions/answers, source statements, internal-link explanations and complete static interface claims/help, while explicitly classifying labels, headings, code and pure download labels. It is not the numerator divided only by the new evidence section. See claim-ledger.csv, denominator-blocks.csv, sentence-audit.json and the independent audit script. The ratio remains an editorial categorization with explicit decisions, not an automated proof of originality.
