Convert a WAV file to MIDI:
Why a short clip can return no notes.
If a very short WAV returns no stable notes in this converter, raising sensitivity cannot create missing analysis windows. Try a longer, clear excerpt containing the intended phrase, then inspect its note draft. Our isolated 0.3-second tone returned zero notes in every mode; the 0.4-second tone returned one in Balanced and More notes. These are measured examples, not a universal minimum clip length.
Same pitch and level. Different clip lengths.
On October 7, 2026, we regenerated five separate mono WAVs and reran this site's unchanged production analysis functions on this Windows machine. Each contains one held A4 sine tone, MIDI pitch 69, with the same amplitude and 8 ms edges. Only the clip duration changes. All regenerated input and MIDI hashes matched the October 6 preparation.
| WAV length | Analysis windows | Cleaner | Balanced | More notes |
|---|---|---|---|---|
| 0.1 s | 0 | 0 | 0 | 0 |
| 0.2 s | 0 | 0 | 0 | 0 |
| 0.3 s | 2 | 0 | 0 | 0 |
| 0.4 s | 4 | 0 | 1 | 1 |
| 0.6 s | 8 | 1 | 1 | 1 |
All nonempty outputs contained pitch 69. A returned event does not establish faithful timing: the 0.4-second source yielded a note about 0.186 seconds long in Balanced and More notes. This experiment diagnoses whether an event survives; it is not a timing-fidelity benchmark.
Why sensitivity sometimes cannot help
This detector analyzes 2,048-sample windows, advancing by 512 samples. For these inputs the downsampling path produces 11,025 samples per second. The inspected implementation uses the following window-count expression, where N is the downsampled sample count:
max(0, floor((N − 2048) / 512))That is this code's actual expression, not a formula assumed for every audio tool. The 0.2-second file supplies 2,205 analysis samples but still produces zero windows under it. The 0.3-second file supplies 3,307 samples and produces two. Moving a sensitivity slider changes acceptance thresholds and the minimum group span; it does not change those sample counts or generate more windows.
Cleaner requires a group span of at least five frame indices; Balanced requires four; More notes requires three. A group span includes the indices from the first accepted frame through the last, not necessarily that many consecutively accepted frames. With only two total windows, our 0.3-second input cannot span three. Four windows allow this particular 0.4-second tone to pass the latter two modes but not Cleaner.
The separate quiet-note sensitivity test concerns a weak note inside a longer recording. Here, lack of analysis context is the issue. A louder source or looser gate is not evidence that enough windows exist.
Choose a retry that tests the suspected cause
- Separate decoding failure from detection failure. A file the browser cannot decode has not reached this analysis stage. A “No stable melody notes were detected” result is a different failure; it does not by itself identify the cause.
- Keep the original and inspect the crop. If you exported only a tiny fragment, retry with a longer actual excerpt around the phrase. Do not overwrite the recording or assume that extending a filename changes the audio.
- Change one thing at a time. Compare the original crop with the longer excerpt at the same setting first. If notes appear, inspect their pitches and boundaries before trying another sensitivity.
- Do not stretch or pad blindly. This run did not test time-stretching, repeated loops or added silence as fixes. Added duration is not automatically additional useful pitch evidence.
- If the result stays empty, check another cause. A stereo source can cancel when averaged; use the channel-sum check for that distinct problem. Overlapping instruments and uncertain pitch require a source-selection check, not this short-clip diagnosis.
A longer excerpt is a diagnostic next step for your own audio, not a guaranteed repair. These five files are complete isolated tones. We did not test a fast short note embedded in a longer song, so the table must not be used to declare such notes undetectable.
Inspect the inputs and outputs
Each WAV is mono, 44,100 Hz, 16-bit integer PCM, with a generated 440 Hz tone at amplitude 0.35 before integer rounding. The MIDI links preserve the unedited writer output for each mode.
- 0.1 s: WAV · Cleaner MIDI · Balanced MIDI · More notes MIDI.
- 0.2 s: WAV · Cleaner MIDI · Balanced MIDI · More notes MIDI.
- 0.3 s: WAV · Cleaner MIDI · Balanced MIDI · More notes MIDI.
- 0.4 s: WAV · Cleaner MIDI · Balanced MIDI · More notes MIDI.
- 0.6 s: WAV · Cleaner MIDI · Balanced MIDI · More notes MIDI.
- Full events, run times, method, source version and SHA-256 hashes.
Method and limits
The repeat regenerated the WAV bytes, parsed their integer samples into Float32 arrays, and called the unchanged downsample, detectNotes and buildMidi functions from the production source. It tested all three modes and checked every generated WAV and MIDI hash against the October 6 originals. Reproduction command in the site project: node tests/wav-short-clip-prep.mjs --check.
No browser audio decoding, file-picker conversion, listening, DAW import, real vocal, instrument performance or full song was measured. This is a code-level test of five synthetic held tones, not a general accuracy percentage, competitor comparison or universal minimum duration. Other implementations may use different windows and grouping rules.
For broad source selection, see the conversion guide. For a downloaded file that will not open, use the MIDI export check. Those questions are separate from whether this detector received enough useful frames to produce a note event.