VOICE → MIDI / Converter Ready

VOICE → MIDI

Voice to MIDI for solo melodies.

LOCAL · PRIVATE

Choose an isolated vocal or humming recording and review the estimated melody notes. Audio stays on this device; one file, up to 5 minutes.

File conversion, not live MIDI. Repeated notes and slides can be missed. See the controlled failure check.

Audio input

Step 01 / 03
Supported audio formats: MP3, WAV, FLAC, OGG, M4A, and AAC. One file, up to five minutes.

Processed locally · never uploaded

Use a clear vocal melody, then check the note events

Voice to MIDI estimates note pitches and timing from a recording. Select an isolated sung phrase, use the converter above, then compare the detected notes with what you sang. The file contains note instructions for an instrument; it does not contain your voice, breath, words or room sound.

For vocals to MIDI, accompaniment creates an additional problem: several pitches can sound at once. This page estimates one pitch at a time and does not separate a singer from a finished mix. Use a solo vocal file when you have one, rather than expecting the MIDI to include every voice and instrument.

A voice melody becomes an editable note estimate: Solo vocal audio, then Pitch estimate, then Review MIDI.
Lyrics and recorded vocal timbre do not travel into the MIDI file.

What to inspect in this particular converter

The note table reports pitch labels, starts and lengths. Check them separately: a matching note name does not prove that an attack or release is correct. The download can pass its structural MIDI check while still needing musical editing. That check verifies the header and track length, not agreement with your singing.

Midilot averages stereo channels into one signal. If a singer occupies one side and another sound occupies the other, the exported file does not preserve them as separate tracks. The MIDI export uses one track, the BPM entered in the tempo field (120 by default), and velocity 88. These settings are not estimates of the singer's tempo or expression.

The detector merges adjacent instances of the same pitch if their gap is shorter than 0.055 seconds. On a repeated syllable, inspect whether two attacks became one longer event. Slides, vibrato and overlapping sounds may need editing because the output is a discrete note estimate, not the complete continuous vocal contour.

Use a known input to learn the checks

The original eight-tone WAV is not a human voice. It was synthesized with programmed C4, D4, E4, G4, A4, G4, E4 and C4 tones, each lasting 0.42 seconds with 0.16 seconds of silence afterward. Its complete duration is 4.64 seconds. It gives you known boundaries before you try a vocal take whose exact ground truth is unavailable.

In the recorded Chrome run on 14 October 2026, the displayed pitch labels followed that sequence. The timing did not consistently follow the input: D4 was programmed to begin at 0.58 seconds, but the note list began it at 0.42 seconds. The displayed difference is 0.42 − 0.58 = −0.16 seconds. This is a reason to inspect timing even after a pitch label looks plausible.

The input parameters, hashes and that run's original rows are available on the WAV reference page. The test demonstrates this detector's behavior on synthesized tones only. It cannot supply a correctness rate for a real singer, whisper, choir or voice memo. For your own recording, compare each phrase against the returned note list before treating it as a playable part.

Try a short clear phrase first. If it has repeated syllables, count the attacks you hear, then count their returned events. If the singer slides between notes, compare the sustained pitch with the surrounding short events. Correct the concrete difference you find instead of assuming the software heard the same musical boundaries you did.

Voice to MIDI free: what is available here

This browser workflow has no account or conversion charge. The same free voice to MIDI route accepts the supported audio formats shown in the picker, when your browser can decode them. The local five-minute limit and single-melody behavior still apply.

Questions about a vocal take

Can I use this page as a live audio-to-MIDI instrument?

This page converts a chosen recording after you select Convert to MIDI. It does not stream live MIDI to a hardware synthesizer. The repeated-tone check below measures a saved file; it includes no live latency measurement. Choose a workflow designed for live tracking if that is your task.

Why did several repeated syllables become one long MIDI note?

Check the attacks rather than the total row count. In our synthetic repeated-tone test, three separated C4 tones became one long C4 event. That is a concrete failure case for this pipeline, not proof about every voice. Inspect the input and downloaded events, then compare the boundaries against your own recording before editing.

Does it convert spoken words into MIDI lyrics?

No. It estimates pitched events and does not transcribe words or produce a lyric track. Ordinary speech is not necessarily a sequence of stable musical notes.

Why does my downloaded MIDI sound different?

Your music software supplies its instrument sound. The MIDI does not preserve the recording's voice or effects. Choose an instrument and review the note timing before arranging the part.

Can I turn the vocal result into a score?

After checking the notes, open the MIDI in the staff viewer. Its notation is a draft of the estimated events, not proof that the singing was transcribed correctly.

A check of the actual voice-page pipeline

On 14 October 2026, choosing the original eight-tone WAV on this page produced eight displayed events and enabled the MIDI download. The D4 row began at 0.42 seconds and lasted 0.58 seconds, while the programmed tone begins at 0.58 seconds and lasts 0.42 seconds. This repeated the detector timing discrepancy on the voice page itself. It verifies a working input-to-output route, not human-vocal accuracy.

Repeated notes and a glide: a controlled failure check

On 16 October 2026, we converted a locally generated two-second WAV on this voice page, downloaded its MIDI, and independently inspected the file with mido. The input is synthetic, not a person singing. It tests whether the current pipeline preserves separated attacks and a changing pitch; it does not establish vocal accuracy.

What the input actually contains

Three C4-reference tones start at 0.00, 0.40 and 0.80 seconds. Each lasts 0.30 seconds, followed by 0.10 seconds of zero-valued samples. A continuous two-semitone glide begins at 1.20 seconds and ends at 2.00 seconds. The frequency starts from 261.625565 Hz; the full sample-by-sample formula and envelope are in the raw record below. Duration is 88,200 samples / 44,100 Hz = 2.00 seconds.

What the downloaded file returned

The 73-byte MIDI contains one track with three note events: C4, C♯4 and D4. In the table, each value is independently read from the download and rounded to six decimal places. The first C4 lasts through the programmed gaps and into the start of the glide. The three repeated C4 attacks are not present as three separate C4 note-on events.

Downloaded MIDI events, seconds
NoteStart (s)Length (s)
C40.0000001.323958
C♯41.3239580.394792
D41.7187500.115625

Swipe or scroll horizontally to view the full timeline →

Three separate 0.30-second C4 tones became one C4 MIDI event lasting about 1.324 seconds.
The gaps in the programmed audio did not become separate C4 attacks in this output.

The source glide changes pitch continuously. The export instead uses note events at fixed pitches, and this file contains no pitch-bend messages. Its last event ends at 1.834375 seconds, before the two-second source ends. There is no single expected discrete-note count for a continuous glide, so counting the three output rows is not a correctness score.

Swipe or scroll horizontally to view the full timeline →

The input rises continuously from a C4 reference toward D4; the MIDI uses C4, C sharp 4 and D4 steps without pitch-bend messages.
Input pitch movement and exported note boundaries are different things to inspect.

How to use this check on a vocal take

Listen for separate syllable attacks, then inspect whether the returned MIDI starts a new event for each attack you intend to keep. Check releases as well as note names. If your phrase contains a slide, decide whether the musical result should use a sustained pitch, discrete notes or an expressive bend in your editor. The raw output from this sample shows why a valid MIDI file can still need that work.

The source code includes a final same-pitch merge rule for event gaps shorter than 0.055 seconds. Our programmed audio gaps are 0.10 seconds. Those numbers refer to different stages: a gap between samples in the input is not necessarily a gap between events before merging. This test records the final file; it does not isolate the stage that removed the repeated attacks. We have not changed the detector as part of this check.

This is one repeatable synthetic case. It is useful for learning what to review, not for predicting the result of a particular singer, microphone or song.