MP3 to MIDI: how audio-to-MIDI conversion actually works

An MP3 contains a recording. A MIDI file contains instructions for playing notes. Converting between them means estimating which notes were played, when they started and how long they lasted. It is closer to transcribing by ear than changing a file extension, which explains both the useful results and the mistakes you will sometimes hear.

Audio and MIDI store different things

Audio formats such as MP3, WAV, FLAC and M4A represent sound: the changing waveform that reaches your speakers. A recording includes the instrument’s tone, room sound, effects and any other sources present at the same time.

MIDI describes musical events. A typical note event has a pitch, a start time, an end time and a velocity value. A MIDI instrument interprets those instructions, so the same notes can trigger a piano, a bass synth or a string patch. The original singer’s voice or guitar tone does not travel inside the MIDI file.

That distinction matters when setting expectations. Audio-to-MIDI can help you recover a melody, sketch a bass line or study a chord progression. It does not recreate a finished recording with all its production choices intact.

The MP3 to MIDI Converter runs Spotify’s open-source Basic Pitch neural note-detection model locally in your browser. Your audio is not uploaded, and you do not need an account. The input can be MP3, WAV, FLAC or M4A, up to three minutes long.

What a neural note detector listens for

A note detector needs to answer several related questions. Is a note beginning? What pitch is present? Is the sound continuing from an earlier note, or is this another strike of the same key?

Short frames describe changing sound

The model works with representations of audio over short, overlapping windows, often called frames. These let it examine how energy at different frequencies changes over time. A note’s fundamental frequency is useful evidence, but its harmonics also matter, especially when the fundamental is weak.

A low guitar note, for example, produces more than one frequency. It has a fundamental plus a series of overtones. A detector must learn that some of those higher frequencies belong to the same note instead of treating each one as a separate key press.

Onsets mark possible note beginnings

An onset is the beginning of a note. A picked string or piano hammer often creates a clear initial transient. A breathy vocal entrance can be much less obvious, particularly when it slides into pitch.

Basic Pitch estimates note onsets alongside continuing note activity. Combining those estimates helps distinguish one sustained note from repeated notes at the same pitch. It is still an estimate: heavy compression, soft attacks and reverb can obscure those boundaries.

Pitch contours follow movement between notes

A pitch contour tracks finer movement than a row of fixed piano keys. That matters for singing, guitar bends and other sounds that glide between pitches.

Basic Pitch can detect pitch bends, so the result may contain information beyond ordinary note numbers. How closely you hear that movement on playback also depends on the receiving instrument’s pitch-bend settings. A simple synth preview is useful for checking the phrase, but it will not reproduce the source instrument’s expression exactly.

The final conversion groups model estimates into discrete MIDI events. Sensitivity and minimum-length controls influence which candidates survive that step. The output is a useful transcription draft, not a guarantee that every detected event belongs in the score.

Monophonic and polyphonic audio

Monophonic audio has one note at a time, such as a solo bass line or an unaccompanied vocal melody. It gives the detector a relatively clear target, even when the tone has many harmonics.

Polyphonic audio contains simultaneous notes, such as a piano chord or strummed guitar. Basic Pitch can estimate overlapping notes, but that does not mean every arrangement is equally easy to transcribe. Notes can mask one another, and their harmonics can overlap.

A solo piano recording can be polyphonic while still being a better input than a full band mix. One instrument has a more consistent sound, and its note attacks may remain clear. A dense mix adds competing instruments, percussion, room sound and effects to the same analysis.

Why full songs are harder

The model does not receive separate labels saying “this frequency is the bass” and “this one is the singer.” A bass harmonic can overlap the melody. A cymbal can create broad energy around several notes. A reverb tail can outlast the note that produced it.

Consequently, a full mix can produce extra notes, missing chord tones and inaccurate note lengths. A busy piano roll is not proof that the conversion captured more music; it may contain more false detections.

If you have access to a clean instrument stem or an original solo recording, use it. The converter’s source presets guide the task, but they do not isolate an instrument from a finished mix.

Choose an input that gives the model a fair chance

For a first attempt, use a short, exposed phrase rather than the most crowded chorus. A dry guitar riff or an unaccompanied vocal line makes it easier to hear whether the output follows the music.

• Start with the cleanest available source. WAV or FLAC avoids additional lossy compression, although a clear MP3 can still work well.
• Avoid clipping. Distorted peaks add frequency content that can confuse pitch estimates.
• Favor dry recordings. Long echoes and reverb can make endings look like continuing notes.
• Keep background sound low. Conversation, drums and room noise compete with the instrument you want.
• Use the original performance speed when possible. Slowing a recording is not an automatic accuracy improvement and can change its tone or attacks.
• Stay within three minutes. For a longer source, prepare a shorter excerpt before selecting it.

Do not spend time renaming a file to WAV or changing its extension. That does not change the underlying recording. Likewise, converting a low-quality MP3 into a lossless container cannot recover missing information.

Convert a phrase and adjust the settings

Open the MP3 to MIDI Converter, select your file and choose the source preset closest to what you want to transcribe: Piano, Guitar, Vocal melody or Bass.

Inspect the piano-roll preview, then use Play MIDI to hear the notes through a simple synth. Compare the contour and rhythm with the source, paying particular attention to repeated notes and long held tones.

The two main controls solve different problems:

Control	If you increase it	What to check
Note sensitivity	More candidate notes may be included	Quiet notes recovered versus extra harmonics or noise
Min note length	Shorter candidate notes are filtered out	Fewer stray notes versus missing quick runs or ornaments

Increase sensitivity when a soft, clearly audible note is missing. Reduce it when the piano roll fills with notes you cannot hear in the original. Raise the minimum note length if tiny fragments litter the result, but lower it again if legitimate short notes disappear.

Change one setting at a time. Otherwise, it is hard to tell whether an improvement came from catching quiet notes or filtering brief errors. Download the .mid when the phrase is close enough to edit efficiently.

Clean up the MIDI in your DAW

FL Studio, Ableton Live, Logic Pro and Reaper can import MIDI for piano-roll editing. GarageBand on Mac can also import MIDI files. Mobile GarageBand workflows and available controls differ, so check your version’s import options before planning a larger edit.

Check the time grid first

Place the imported phrase against a suitable tempo and listen to where the beats fall. Audio-to-MIDI note timing is not the same thing as a reliable tempo map. Before quantizing, make sure the DAW’s grid corresponds to the performance you want to preserve.

Remove wrong notes before moving good ones

Delete obvious outliers above or below the part. Then check octave errors, overlapping duplicates and notes that continue through a rest. For a bass line, listen for false higher notes that may actually be harmonics of the bass.

Loop a short section and work from the clearest notes outward. A convincing four-bar phrase is more useful than a long import with unexamined errors.

Quantize with restraint

Quantizing moves note starts toward a rhythmic grid. Choose a grid that fits the phrase: a sixteenth-note grid can be wrong for triplets, and a rigid grid can flatten a loose performance.

Use partial quantization or move individual notes when the original groove matters. Correct note lengths separately, especially where a sustain pedal or reverb has blurred the recording’s endings.

Adjust velocity and pitch bends

MIDI velocity influences how strongly an instrument plays a note; it is not a direct copy of recorded loudness. Set a sensible range for the chosen patch, then shape accents by ear.

Check pitch bends with the instrument you intend to use. If a bend sounds too wide or too narrow, the receiving instrument’s bend range may need adjustment. Remove accidental bends only after comparing them with the original performance.

Personal study and publishing are different uses

Transcribing music can be useful for ear training, but converting a copyrighted song does not grant permission to distribute its melody or publish a derivative arrangement. Private study and public release are treated differently across jurisdictions; personal use is not a universal legal exemption.

For material you plan to publish, check the rights to the underlying composition and any recording you reuse. Your own performances, public-domain compositions with appropriate recordings, and suitably licensed material provide clearer starting points.

Once your notes are ready, you can explore the other music tools for a separate audio-editing task. If the finished audio needs a readable title or embedded artwork, the guide to MP3 tags explains that final library step.