# Change MP3 to MIDI — source-history control, 2026-11-11

Convert a clear single-note input with the existing browser tool, inspect its notes and download MIDI. This page evaluates a concrete source-choice question: does an existing MP3 decoded to WAV, or resaved at a higher bitrate, recreate the draft obtained from the original recording? It does not claim to restore an unknown missing source. The unchanged algorithm detects one pitch at a time; simultaneous polyphonic transcription requires another processing path.

## Recorded source and predeclared controls

Xi and colleagues’ GuitarSet1.1.0, DOI https://doi.org/10.5281/zenodo.3371780, CC BY4.0. See ATTRIBUTION.md, reference.json and source-check.json. Real human acoustic-guitar microphone recording, not rendered MIDI or a synthetic instrument. Full already-public recording and annotation:

- https://audiotomidinow.com/guitar-to-midi/files/00_BN1-129-Eb_solo_mic.wav
- https://audiotomidinow.com/guitar-to-midi/files/00_BN1-129-Eb_solo.jams

Original file SHA256 `2a6057d0485d8ee2f4a04175ee085d610c0f7a13194e083b13ce803bec091d43`; JAMS SHA256 `2258ef6f5dc58437f2807dfc0c05416ad8498cd345f95d6563b13d19283e6d54`. Four adjacent windows [0,5), [5,10), [10,15), [15,20) seconds were fixed before conversion, not selected for outcomes. The complete performance is about22.3244seconds; the remaining tail is outside this panel. A/B/C/D are correlated excerpts of one performance, not four independent subjects. Input selection/method is not counted as original evidence.

Each source crop is44100Hz mono signed16bit PCM,220500samples. Seven paths per window give28 offline conversions at unchanged Balanced mode2:

1. Original WAV.
2. Original →192kb/s MP3 →PCM16 WAV (`direct192`).
3. Original →32kb/s MP3 →PCM16 WAV (`low32-g1`).
4. Existing low32 MP3 →FFmpeg stream copy with changed ID3 title →PCM16 WAV (`low32-retag-copy`).
5. Decoded first low32 →192kb/s MP3 →PCM16 WAV (`low32-to192`).
6. Decoded first low32 →32kb/s MP3 →PCM16 WAV (`low32-g2`).
7. Decoded second low32 →32kb/s MP3 →PCM16 WAV (`low32-g3`).

No normalization, mode tuning, source retiming, detector change or synthesized audio. Actual encoder commands are in generate.py; actual version and binary SHA256 in inputs.json. The archived FFmpeg executable hash is `c8abc49e7be62dde8e12972af373959e0076a7b8dc8040eb45978e0608f8781e`. Official source tag n4.1 is only the documented frame-parser cross-check, not an assertion that the installed build exactly matches that tag. Different encoder/decoder builds can produce different bytes and results.

## Independent verification

`mp3_frames.py` parses this limited MPEG1LayerIII framing, excludes Info/Xing metadata frames from the coded-audio hash, and checks header bitrate/rate. It is not a separate audio decoder. Container, complete MPEG frame, coded audio, decoded PCM and MIDI byte equality are separate tests. File hashes alone do not establish waveform or note equality.

`midi_reader.py` parses stored events independently of the JavaScript writer. `check.py` compares those events with raw detector output, makes pitch/time CSV and generated Piano Roll SVG for all28 conditions, and matches inherited reference attacks to nearest unused equal-integer-pitch outputs within150ms in reference chronological order, one-to-one, without a duration gate or octave equivalence. Source notes already active at a crop boundary are included in reference plots but excluded from the new-attack denominator. Reference intervals can overlap because guitar strings ring. Labels are largely automated, not freshly adjudicated listening truth; no population accuracy score is claimed.

Parsed format0 files have one track,480PPQ and500000microseconds/quarter, so960ticks/second at that tempo. Start and duration values are parsed seconds; ms =1000×seconds. Velocity is an integer MIDI event field, not measured perceived loudness. `compare.py` computes exact per-event arithmetic, PCM RMS difference to the original on a full-scale basis, packet/PCM/MIDI identity and the source-history comparisons. A pitch/onset match does not establish duration fidelity. A lower RMS error did not monotonically order the selected detector match count; that is a limited counterexample, not a codec or perceptual quality ranking.

## Browser controls and reproducibility

On2026-11-11, four actual production file-chooser conversions/downloads used A/D low32-g1 MP3 and their respective decoded PCM16 WAVs, Balanced2. All four saved MIDI files are unmodified and byte-identical to their corresponding harness results. The MP3/WAV pair is identical in each window. Only these four UI operations were tested; the remaining numerical runs are offline controls. No network-privacy, listening, DAW-import or competitor comparison is inferred.

Download all files in manifest.json into one directory. The four source crops, reference, scripts, detector snapshot and Python MIDI parser are required; browser comparison additionally needs the four archived browser MIDI files. With Python3, Node and a trusted FFmpeg build supporting libmp3lame:

```text
python source-check.py
python generate.py --ffmpeg /path/to/ffmpeg
node convert.mjs
python check.py
python compare.py
python browser-compare.py
```

The source check can instead use `--original /path/to/original.wav --annotation /path/to/reference.jams`; otherwise it downloads the two existing public files above and verifies their hashes before cropping/reconstructing references. A fresh-directory rerun with the archived identical build reproduced all140 numerical artifacts enumerated in reproduction.json. The browser files are archived fixed inputs, not regenerated or counted in140. Full original audio and JAMS are existing downloads, not newly generated evidence. `claim-ledger.csv`, `denominator-blocks.csv`, `sentence-audit.json` and `claim-bindings.json` disclose the complete-main sentence denominator, exclusions, source scope and independent claim checks.

## What changed and what did not

The September24 synthetic four-note dataset remains unchanged at `/assets/change-mp3-to-midi/`; it is not counted as today’s recording or original research. November10’s retained-PCM silence/context experiment and the homepage’s mode/channel baselines are separate controls. The new source-history question is about packet-copy identity, equal-final-bitrate histories and repeated lossy generations, not another report of the old synthetic64/192 result. Production algorithm and site structure remain unchanged.

The Google snapshot is dated2026-11-11 with hl=en/gl=us/pws=0; physical country was not independently verified. Ten complete readable bodies were reviewed, with the first three intact; original rank5 Reddit returned403 and remains unknown, and observed-order rank11 provided the tenth readable body. Only the actually retrieved public bodies support the limited absence statement, not linked/hidden/unavailable content or worldwide novelty. Primary standards, setup, input descriptions, inherited results, scope and necessary instructions are excluded from originality credit.
