Everyone has hummed a tune that felt like it could be something. Usually it evaporates by lunchtime. This project catches one and carries it all the way to a short finished piece: a melody, chords under it, two sections and a bounce you can send to someone.
Here's the moment that makes it possible. Press play: this is what a hummed tune looks like after a model turns it into notes. It's nearly right. One note isn't.
The last note sounds sour. Move it by a semitone or two until the phrase lands. Which note did the singer mean?
Fix what the model heard
Loading the phrase…
The last note sounds sour, like the tune trips on the final step. Move it up one semitone, to G. Now the phrase settles, and it sounds like the tune somebody actually meant. The model supplied a guess; you corrected it by ear. Before you start: what do you think a notes-only transcription will lose from your hum, beyond the odd wrong note?
From hum to piece
Stage 1: Catch the idea
Hum the tune into your phone. Keep it under fifteen seconds, one clear line with a pause at the end. Don't sing it beautifully; sing it clearly. Then listen back once and write a single sentence about what makes it worth keeping: a leap in the middle, a rhythm, the way the last note sits.
That sentence is your compass. Every later decision should protect it.
Checkpoint: a saved recording and one sentence about its identity. Melody and motifs helps you hear what's memorable in a short phrase.
Stage 2: Turn it into notes
Run the recording through an audio-to-MIDI tool: give it the clip, let it estimate the notes, then export them as a MIDI file. Choose a tool that accepts your recording and lets you bring the result into your MIDI editor.
Look at what came back, and listen to it next to your hum. A pitch model listens for the strongest, lowest repeating line in the sound. Your voice has many lines stacked above the note you're singing:
Find the lowest bright line in the spectrogram. Can you see the fainter copies above it, and why a model might sometimes pick the wrong one?
What the model has to read
ReplaySpectrogram window: 2048 samples, 42.7 ms at 48 kHz; bins every 23.4 Hz. Shown up to 12 kHz.
That's why the usual mistakes are notes an octave too high, tiny extra notes where your voice wobbled, and notes split in two where you took a breath. Now you know what to look for.
Checkpoint: a MIDI file of the estimate, kept as it came out. Audio to MIDI explains how the model listens and where it slips.
Stage 3: Fix what changes the tune
Open the MIDI in any editor with a piano roll. Play it next to your hum, a phrase at a time, and fix only what changes the tune: a wrong pitch, an octave jump that shouldn't be there, a stray blip. Keep a copy of the raw estimate so you can see what you changed.
Don't snap everything to the grid. A note that lands a little late might be the very thing your compass sentence was about. Quantize only where a steady pulse makes the phrase better, and listen after each change.
The file won't carry your slides, your breath or the color of your voice. Notes are only where and when. You'll add expression back later with the instrument you choose. And the tempo written in the file is the tool's default, not the speed you hummed at; set your own.
Checkpoint: a corrected melody that, played on a piano sound, still feels like your hum. Rhythm, meter and groove helps you decide where the timing should be strict and where it shouldn't.
Stage 4: Put chords under it
Listen to the melody's long notes and its last note. They tell you a lot about which chords fit. Start with a bass note and a few simple chords, held, one per bar.
Switch between root position and voice-led. Which version leaves more room for a melody on top?
Chords that move gently
Loading the phrase…
Voice the chords so their notes move as little as possible from one to the next, like the voice-led version above. Now the harmony glides under your melody instead of jumping around it, and the melody stays the thing people follow.
If a melody note clashes with a chord, decide whether the tension sounds good or wrong. Sometimes a rub is the best moment in the piece. If it's wrong, change the chord, not the tune. Stuck for options? A chord-suggestion model can offer a few to choose from, and Generating notes explains why its ideas swing between cliché and chaos, and how to steer them toward your compass sentence.
Checkpoint: melody, bass and chords playing together, with one sentence on why your favorite chord change works. Chords and voicing goes deeper.
Stage 5: Grow it into two sections
A loop isn't a piece yet. Make a first section that states the tune plainly: melody and held chords, eight bars. Then a second section that changes one thing: the melody an octave up, the chords played in a rhythm, the bass moving more, or the tune's last phrase answered by a new one.
Play from start to finish without stopping. Does the second section feel like somewhere new, or like the first one, only louder? If it's only louder, take something away instead of adding.
Checkpoint: two sections, sixteen bars or so, that play through. Form and arrangement shows how contrast makes a section boundary feel intentional.
Stage 6: Choose the sounds
Pick a sound for the melody that suits your hum: a soft electric piano for a gentle tune, a plucked sound for a bouncy one, a sampled instrument you made yourself if you've done the first sampled instrument project. Test its highest and lowest notes in your tune; some instruments sound thin at the top or muddy at the bottom.
This is where expression comes back. A little vibrato, a slide on the last note, a longer release: things the transcription couldn't carry, now added on purpose. Now your melody sings again.
Checkpoint: every part has a sound, and the piece plays through. Composing with a library has more on fitting parts to an instrument's strengths.
Stage 7: Finish and share it
Balance the parts so the melody is clearly in front. Listen from start to finish, including the tail at the end. Fix the one thing that bothers you most, then stop. Bounce it.
Now play your original hum and then the bounce, back to back, for a friend. Hearing a fifteen-second phone recording turn into a finished piece is genuinely delightful, and it's the best proof that the idea was worth keeping.
Checkpoint: a bounce, with the original hum, the raw estimate and the corrected MIDI kept beside it.
Quick check
- The model's melody jumps up an octave on one note. Is that what you sang? Probably not. Pitch models sometimes pick the harmonic an octave above a weak note. Trust your ear and the recording, and move it back down.
- After quantizing, your melody sounds stiff and loses its charm. What happened? The grid removed timing that was part of the tune. Undo it, and quantize only the notes that sound wrong off the grid.
- The same melody works over C Am F G and over C Em F G. Which is correct? Both. A melody supports several harmonizations; each gives it a different mood. Choose the one that matches your compass sentence.
Next steps
- Use it: Bass lines and counterlines. Give the bass a line that supports the melody without copying it.
- Finish it: Finishing and references. Check your short piece against a reference at a fair listening level.













