Audio to MIDI converter:
check the notes you keep

Audio to MIDI turns a recording into editable note events; this converter follows one pitch at a time, so use an isolated part and check its saved draft, while simultaneous parts need a polyphonic transcription path. Evidence

Open the browser converter · Compare the five tests · Download the recordings and MIDI

1. What this comparison measures

All five actual browser downloads matched the unchanged production algorithm’s offline MIDI byte for byte, allowing the saved event comparisons below without claiming a general converter accuracy rate. Evidence

The batch reuses one human GuitarSet recording, applies declared processing controls and adds a real Vienna piano performance for a controlled two-instrument mix, without replacing either instrument with synthesized notes. Evidence

The frozen main match requires the same integer MIDI pitch and an attack within fifty milliseconds, choosing the nearest unused output in reference order without fitting a shift or crediting an octave substitute. Evidence

Reference attacks come from per-string guitar annotations and calibrated piano hardware key events, so overlapping strings, pedal decay, separate recording rooms and calibration uncertainty limit what this panel can establish. Evidence

The saved MIDI uses format zero with one track at 480 ticks per quarter note and separate Note On and Note Off events, while the accompanying time tables use seconds and integer MIDI pitch numbers. Evidence

2. Five conditions, with the failures visible

Condition	Reference attacks	Output events	Strict matches	Missed	Extra	Reference / output maximum simultaneous notes
Real plucked recording	20	8	0	20	8	3 / 1
Controlled vibrato on real recording	20	8	0	20	8	3 / 1
Declared delay/reverb control	20	8	0	20	8	3 / 1
Added white noise	20	8	1	19	7	3 / 1
Controlled real guitar + piano mix	45	10	0	45	10	6 / 1

In this fixed batch every output stayed at one simultaneous note, including the mix whose reference reached six, so retaining one melodic line must not be presented as recovering its simultaneous parts. Evidence

Plucked notes: pitch identity and attacks are separate checks

The plucked recording produced eight events against twenty annotated attacks and zero strict matches, leaving twenty missed references and eight unmatched outputs under the frozen rule. Evidence

Five of those eight events nevertheless overlapped a same-pitch reference interval, so the zero strict score cannot be read as every stored pitch being wrong. Evidence

Three remaining plucked outputs overlapped only octave-related reference pitches, which identifies concrete candidates for manual register checking without awarding them exact-pitch credit. Evidence

A posthoc wider attack window yielded three plucked matches instead of zero, demonstrating that the declared timing tolerance changes this count without repairing the MIDI itself. Evidence

Vibrato: a stable pitch list does not prove expressive recovery

The controlled-vibrato output retained the baseline sequence 52, 57, 33, 57, 38, 62, 58 and 34, but its complete event data changed, so matching pitch lists alone would hide a changed draft. Evidence

In the vibrato draft the third output, pitch 33, began about 46 milliseconds later while the sixth, pitch 62, began 325 milliseconds earlier than their baseline counterparts, so one global time offset cannot align both exported drafts. Evidence

The vibrato MIDI contains zero Pitch Bend channel messages, which rules out treating this particular saved file as evidence that continuous expressive bends were recovered. Evidence

Vibrato and plucked both scored zero with the primary timing rule, yet the wider exploratory window accepted four and three respectively, exposing timing sensitivity that identical event totals conceal. Evidence

Reverberation: equal totals can hide a register change

Reverberation kept the total at eight events while same-pitch interval overlaps fell from five to four and octave-only overlaps rose from three to four, so the equal event total concealed a different pitch allocation. Evidence

The sixth sorted output changed from pitch 62 to 38 and began about 140 milliseconds earlier in the exported MIDI, an ordinal two-octave difference that does not by itself prove both events identify the same performed note. Evidence

White noise: a boundary crossing is not an accuracy promise

Added noise preserved the baseline pitch sequence but moved one pitch-58 attack inside the frozen timing boundary, changing strict matches from zero to one without establishing that noise improves transcription. Evidence

That accepted noisy attack arrived about 47 milliseconds early and its stored duration was about 587 milliseconds shorter than the reference, showing that an accepted onset still needs an ending check. Evidence

Controlled mix: more source notes did not become more simultaneous parts

Adding the real piano increased reference attacks from twenty to forty-five but output events only from eight to ten, so this mix did not turn the single-line draft into a transcription of the added part. Evidence

Six mix outputs shared intervals with exact-pitch references while four had only octave-related overlaps, so zero strict matches must again be separated from complete loss of pitch relationships. Evidence

The mix recovered three strict-pitch attacks only after the exploratory timing window was widened, while maximum output polyphony remained one, separating an onset criterion from the simultaneous-note limit. Evidence

3. Read the diagnostics without moving the goalposts

Interval-overlap counts and the wider 150-millisecond window were added after viewing the main outputs, count outputs differently and must not be substituted for the frozen fifty-millisecond headline matches. Evidence

The five-condition comparison therefore supports inspecting register, attacks, endings and simultaneous notes separately, rather than labeling a draft correct solely because an event count or pitch list looks stable. Evidence

4. Listen, inspect the Piano Rolls and reproduce

The SVG Piano Rolls are generated reference-versus-output plots rather than DAW screenshots, and the time tables describe MIDI events verified against the actual browser downloads. Evidence

An independent fresh run regenerated fifty common audio, MIDI, reference, table, plot and numeric-result files with identical SHA256 hashes, proving reproducibility of this offline pipeline rather than browser compatibility. Evidence

Frozen protocol · Source transforms · Primary results · Posthoc diagnostics · Unchanged detector · Offline conversion runner · Independent MIDI parser · Source licenses and attribution · Fresh-run SHA256 checks

Reproduction instructions · Portable count and plot scorer · Posthoc diagnostic script · Original Windows input generator · Independent prose arithmetic · Staged evidence SHA256

Five actual browser downloads · Named vibrato timing comparison

Internal evidence sentence ledger · Internal full-body audit · Dated source review and limits · Recording provenance checks

5. Supported formats and browser verification

The ten fixtures were actually tried in this Windows Chrome session: MP3, WAV, FLAC, OGG, M4A, AAC, MP4, MOV, WebM reached a downloaded MIDI, while AIFF failed; these results do not guarantee every codec profile or browser. Evidence

Tested fixture	Actual browser result	Saved events / evidence
MP3	Decoded, converted and downloaded	7 MIDI events
WAV	Decoded, converted and downloaded	8 MIDI events
FLAC	Decoded, converted and downloaded	8 MIDI events
OGG	Decoded, converted and downloaded	8 MIDI events
AIFF	Decode failed in this browser	Actual error and input SHA256
M4A	Decoded, converted and downloaded	8 MIDI events
AAC	Decoded, converted and downloaded	9 MIDI events
MP4	Decoded, converted and downloaded	8 MIDI events
MOV	Decoded, converted and downloaded	8 MIDI events
WebM	Decoded, converted and downloaded	9 MIDI events

Why did the AIFF file fail?

This AIFF sample stopped before note detection with “Unable to decode audio data” in the tested browser; export an ordinary PCM WAV in your audio editor as the next input to try, rather than treating AIFF as supported here. Evidence

6. What the controls do and do not represent

Vibrato is a declared five-hertz periodic time warp of the real recording, reverb is an explicit three-tap delay treatment, noise uses a fixed seed and the mix sums two real performances instead of representing a commercial full-band recording. Evidence

Every audio fixture includes the same half-second tail and contains no clipped PCM samples, while the guitar reference itself contains overlapping string annotations and must not be relabeled a strictly monophonic ground truth. Evidence

GuitarSet and the Vienna piano materials are credited under CC BY 4.0, with our crops, effects, alignment, note analysis and plots disclosed as modifications without source-author endorsement. Evidence

7. Local processing and the next check

For the observed WebM selection and MIDI generation, Chrome’s unfiltered Network list remained empty and the converter read the local file into a local MIDI blob; this is a scoped conversion check rather than a claim that initial site assets or analytics never connect. Evidence

Network before file selection · Generated MIDI and empty network list · Actual WebM result · Observation boundaries · Unchanged production client source

Start with a short isolated part, keep the recording and reference available, inspect an unwanted register or merged attack in the saved events, and use a polyphonic processing path when simultaneous parts are required. Evidence

Guitar recording checks · Piano recording checks · WAV and MIDI directions · MIDI channels