What we tested
We generated one mono WAV file at 22,050 Hz containing five isolated sine tones. C4 lasted 30 milliseconds, D4 lasted 50 milliseconds, E4 lasted 80 milliseconds, F4 lasted 120 milliseconds, and G4 lasted 200 milliseconds. Each tone was followed by 350 milliseconds of silence. No public recording, video, performer, or channel was used.
We ran the file through the current local converter at its default 0.25 sensitivity. First we kept the default 0.06-second minimum note length, then changed only that control to its lowest available value of 0.02 seconds. The browser completed without console errors in both cases.
| Input | Default 0.06 s | Minimum 0.02 s |
|---|---|---|
| C4 · 30 ms | Not exported | Not exported |
| D4 · 50 ms | One D4 · 93 ms | Two D4 events · 47 ms each |
| E4 · 80 ms | One E4 · 104 ms | Two E4 events · 70 and 34 ms |
| F4 · 120 ms | One F4 · 140 ms | Two F4 events · 104 and 34 ms |
| G4 · 200 ms | One G4 · 222 ms | One G4 · 222 ms |
What the minimum note length control actually changed
The lower setting admitted shorter detected segments, but it did not create evidence the model had not found. That is why the 30 millisecond C4 stayed absent. It also turned the D4, E4, and F4 detections into adjacent fragments, increasing the MIDI from four notes and 87 bytes to seven events and 110 bytes.
The control is therefore a cleanup tradeoff, not a promise that every source note longer than the chosen number will appear. Source attack, pitch, level, timbre, overlap, and the model's frame-by-frame confidence can all affect the draft.
How to handle short articulations
- Start with the default minimum. Check whether the important attacks are present before changing a control.
- Lower the minimum one step at a time. Compare the new attacks with the source and watch for adjacent duplicate fragments.
- Keep the cleanest draft. More MIDI events do not automatically mean a closer transcription.
- Add a missed note by ear. For a clearly audible attack that never appears, drawing one note can be faster than accepting many false fragments.
What this test proves—and what it does not
It records the exact output for five generated mid-register sine tones in the current converter. It does not establish a universal cutoff for staccato piano, muted guitar, percussion, vocals, compressed audio, or a dense mix.
If you see too many adjacent events, use the extra-note cleanup guide. If repeated attacks merge together, read the repeated-note test. For timing against a beat grid, see the tempo guide.