WAV to MIDI: silence gives 0 events; a noisy tail does not add notes in this test.
A silent WAV produced zero MIDI events. A four-note melody followed by a measured low-level noise tail still produced four events, with the last note ending before the tail. Download both inputs and outputs to check the result yourself.
The four files in one view
| Input | Output | Events |
|---|---|---|
| silence.wav | silence.mid | 0; “No clear notes found” |
| melody-with-noise-tail.wav | melody-with-noise-tail.mid | 4: MIDI 60, 64, 67, 72 |
The second file is 3.8 seconds long. Four 0.5-second tones end at 2.6 seconds; random noise at amplitude about 0.035 begins at 2.65 seconds. The detector returned four events and did not turn the tail into an extra note.
What this does and does not prove
This is a controlled code-path test, not a universal noise threshold. The random tail is synthetic, mono PCM at 11,025 Hz and 16-bit quantization. A real recording can contain louder noise, hum, breaths or a note that fades into the tail. Check the final note boundary in the MIDI instead of assuming the tail was ignored.
Reproduce the test
- Download the two WAV files and their MIDI outputs.
- Download run-noise-tail.cjs, which generates the inputs and evaluates the deployed detector.
- Read the complete results.json, including event counts, note boundaries and the tested app source hash.
WAV to MIDI questions
What happens when the WAV is silent?
The converter reports no clear notes and emits zero note events in this test. A valid WAV container alone does not guarantee a MIDI download.
Can a noise tail create a false note?
Not in this measured tail: the four known notes remained the only events. Treat that as evidence for this amplitude and implementation, not as a guarantee for every recording.
What should I inspect after conversion?
Compare the last detected note’s end time with the source audio. If the tail contains a real note or a louder artifact, lower confidence or trim the source and repeat the check.