Voice to MIDI

A control-by-control workflow for this tool

Choose your file and wait for the audio-loaded message. Use the input audio player to check that you selected the intended take. The source waveform and duration belong to your input; the piano roll and note count below it belong to the MIDI result. Keep those two displays separate when deciding what needs fixing.

1. Set Lowest note and Highest note before conversion. These controls restrict the pitch search. They are not a singer classification menu. Choose a range that contains the part you want, then press Convert to MIDI. If you change either control after a result appears, the site displays a request to convert again. The previous piano roll remains visible until you do so; it is not evidence that the new range has already been analyzed.
2. Review the result at Transpose 0. Use Preview MIDI to hear the detected notes through the built-in synth. Compare the melody with the input player, which plays the recording. The two sounds will have different timbres. Transpose moves the detected notes together; it cannot isolate one wrong note or turn a backing track into a vocal stem.
3. Use Shortest note for short unwanted fragments. Increasing this value removes detected note segments shorter than the selected duration. It does not join all fragments into one sustained note. If an intended short syllable disappears too, reduce the setting and compare again. This control updates an existing conversion without rerunning pitch analysis.
4. Use Quiet-sound filter to change the level threshold. Moving toward a less negative value makes the threshold stricter. That can also remove quiet intended notes. Change one control at a time, so you can tell whether a missing note followed the level adjustment or a different setting. This is a filter on the analysis, not a noise-removal effect applied to your original recording.
5. Adjust Detection confidence separately. A higher threshold keeps fewer low-confidence detections. The displayed value is a filtering threshold, not an accuracy percentage. If increasing it leaves gaps, compare with the earlier setting. Neither this slider nor the level filter can recover a melody that is absent from the recording.
6. Download after your final adjustment. Transpose and the three filters refresh the current MIDI download. Save the file after reviewing those changes. Changing the pitch-search range is different: run Convert to MIDI again first. Reloading or closing the converter clears its current result, so it is not a saved project.

What we actually verified

On September 22, 2026, we used the live site's built-in synthetic sample to check the controls in desktop Chrome. This was an interface check, not a test of a human singer. With the default settings, the display showed 8 notes and the range C4–A4. Moving Transpose from 0 to +12 changed the range to C5–A5 while keeping 8 notes, without another conversion.

Next, we moved Lowest note from C3 to C-sharp 3. The site displayed “Pitch range changed. Convert again to apply it.” The old result still showed C5–A5. That is why the workflow above separates pitch-search settings from controls that update an existing result immediately. You can repeat this check with Try a sample melody before selecting your own vocal recording.

The test confirms those controls and messages for this example. It does not establish transcription quality for singing, vibrato, breathy passages or layered voices. We have not used this synthetic sample to claim human-vocal accuracy.