MP3 to MIDI note-boundary check: compare event edges with known controls.
In this controlled WAV run, the four-note fixture produced four events, the five-note fixture produced five, and the silent control produced zero.
Input condition
Generated mono PCM WAV fixtures were run through the exact deployed detector and MIDI writer in Node VM. The recorded outputs let you compare event count and note start/end boundaries without treating a control file as a full-song benchmark.
| Input | Observed output | What to inspect |
|---|---|---|
| piano-like-melody.wav | Four events: pitches 60, 64, 67, 72 | Four sequential note boundaries |
| five-note-melody.wav | Five events: pitches 55, 59, 62, 67, 71 | Whether the extra note is retained |
| silent-control.wav | Zero events | Quiet-input behavior |
Outputs and the complete results.json remain available with the fixtures.
What this does not prove
This control run does not establish full-song accuracy, polyphonic coverage, or performance on every recording. It documents only the generated fixtures, detector, and recorded boundaries.
After conversion
Compare event count first, then inspect each note start and end against the expected fixture. Keep the input, MIDI output, and result record together.
Questions about note boundaries
How many events did the four-note control produce?
It produced four recorded events with pitches 60, 64, 67, and 72.
What does the silent control tell me?
It produced zero events under the same detector run.
Can these controls prove a full recording will match?
No. They are reproducible controls for checking event boundaries, not a universal accuracy test.