Audio to MIDI velocity:
inspect notes before volume
Audio-to-MIDI velocity gives you editable strength values for individual notes; inspect those values and the source-checked notes before changing track volume. velocity-analysis.json · primary-sources.json
The comparisons below are a November 17 analysis of unchanged browser exports captured on November 16, 2026, with no new recording, conversion or model run. source-index.json · browser-verification.json
2. Read the note-strength field
In MIDI 1.0, Note On carries a key number and a separate seven-bit velocity value, while channel volume uses a Control Change message. primary-sources.json · summary-of-midi-1-0-messages
The inspected worker rounds model amplitude multiplied by 127 and limits its exported velocity to 1–127; that rule is not a decibel measurement. worker-velocity-rule.txt
3. Comparison 1: variation is not correctness
The saved guitar draft contains 55 notes with 34 distinct velocities from 28 to 109 and a median of 74. velocity-analysis.json · note-velocity-ledger.csv
The piano draft also contains 55 notes, but has 40 distinct velocities from 29 to 107 and a median of 73. velocity-analysis.json · note-velocity-ledger.csv
The direct mixture has 100 exported events and 51 distinct velocities, with its median down at 53. velocity-analysis.json · note-velocity-ledger.csv
Moving the piano component later in the archived timing control gives a 101-event draft with 55 distinct velocities and a median of 59. velocity-analysis.json · source-index.json
These distributions include every stored event, including unmatched outputs, so variation alone does not establish faithful musical dynamics. note-velocity-ledger.csv · velocity-analysis.json
In the guitar ledger, matched outputs have median velocity 88.5 while unmatched outputs have median 65, with their ranges overlapping from 83 through 94. velocity-analysis.json · note-velocity-ledger.csv
The piano split has medians of 82 for matched outputs and 59.5 for unmatched outputs, with the unmatched range extending up to 104. velocity-analysis.json · note-velocity-ledger.csv
Among piano events with velocity at least 90, eight are matched and seven are unmatched in the existing reference ledger. velocity-analysis.json · note-velocity-ledger.csv
That descriptive cutoff is not a deletion filter, and an unmatched event is not automatically an invented pitch because the crop and annotation have stated limits. velocity-analysis.json · protocol.json
Matched labels are reused unchanged from the original exact-pitch, 50-millisecond-onset ledger, rather than reassigned using velocity. protocol.json · note-velocity-ledger.csv
| Archived input | Events | Velocity range | Distinct values | Median |
|---|---|---|---|---|
| guitar | 55 | 28–109 | 34 | 74 |
| piano | 55 | 29–107 | 40 | 73 |
| mix | 100 | 26–103 | 51 | 53 |
| offset-mix | 101 | 28–105 | 55 | 59 |
4. Comparison 2: the same referenced guitar attacks
Sixteen guitar reference attacks already matched in both isolated and mixed drafts link to twelve lower, two equal and two higher stored velocities in the mixture. velocity-analysis.json · guitar-context-velocity.csv
Across those linked attacks, mixture-minus-isolated velocity ranges from −31 to +9, with a median difference of −4.5. velocity-analysis.json · guitar-context-velocity.csv
For example, the linked pitch-55 attack beginning near source time 10.054780045 seconds changes from velocity 95 to 89, while the next pitch-67 attack changes from 109 to 103. guitar-context-velocity.csv
Those links use known reference-source labels and original attack times; they do not mean that the model identified a guitar track or isolated the cause of a velocity change. guitar-context-velocity.csv · ref-guitar.json · ref-mix.json
| Linked reference attack | Solo velocity | Mixture velocity | Difference |
|---|---|---|---|
| Guitar pitch55 @10.054780045s | 95 | 89 | -6 |
| Guitar pitch67 @10.278385488s | 109 | 103 | -6 |
| Guitar pitch55 @10.481129252s | 100 | 86 | -14 |
5. Comparison 3: exported and reference piano velocities
For the 21 archived piano matches, fifteen exported velocities are lower than the reference MIDI values, two are equal and four are higher. velocity-analysis.json · piano-reference-velocity.csv
Export-minus-reference differences range from −42 to +19 raw velocity units, with a median absolute difference of 15. velocity-analysis.json · piano-reference-velocity.csv
Among 204 pairs whose reference velocities differ, the exported order agrees for 118 pairs, reverses for 80 and becomes a tie for six. velocity-analysis.json · piano-pair-ordering.csv
Pairs tied in the reference are excluded from that ordering comparison, and these descriptive counts do not calibrate model amplitude to recorded key-strike velocity or acoustic loudness. piano-pair-ordering.csv · piano-source-provenance.json · worker-velocity-rule.txt
| Piano reference attack | Reference velocity | Export velocity | Raw difference |
|---|---|---|---|
| Pitch71 @0.445s | 114 | 100 | -14 |
| Pitch64 @0.45625s | 94 | 94 | 0 |
| Pitch56 @0.46625s | 98 | 83 | -15 |
| Pitch64 @0.94875s | 97 | 101 | 4 |
6. Edit notes before the overall level
A proposed flatten-to-64 edit would overwrite all 55 guitar values and 54 of the 55 piano values, based only on counting existing values unequal to 64. velocity-analysis.json · note-velocity-ledger.csv
Keep the raw draft, correct source-checked pitch and timing problems first, then edit individual velocity values where the intended emphasis needs it. note-velocity-ledger.csv · piano-reference-velocity.csv
Channel volume and expression are separate controls, so inspect note velocities separately from those controller events and the receiving instrument’s response. primary-sources.json · note-velocity-ledger.csv · midi-1-0-control-change-messages
The older piano browser/offline comparison still differs at byte 122 in one velocity value, 56 versus 55, and neither file has been replaced or described as byte-identical. runtime-comparison.json
7. Dates, method and limits
The recordings, crop alignment, mixed-source construction and annotation uncertainties remain those of the original study, and no new listening, instrument response or general conversion-accuracy benchmark is claimed. ATTRIBUTION.md · protocol.json · worker-velocity-rule.txt
The included analysis script only reads the saved CSV, JSON and MIDI files to rebuild these tables, without decoding audio, running a model or requesting another browser conversion. analyze.py · source-index.json
The source comparison reuses a dated core-family review, not a current Google ranking for this phrase or a claim of exclusive functionality. research-review.json
8. Continue with the relevant check
Use the converter when you need a new note draft, retain its raw MIDI, and use these archived examples to decide which note fields to inspect before editing. source-index.json