Controlling Melody
Humming gives contour, MIDI gives notes, and an audio reference lends its style along with the tune.
Related
Three inputs, three outcomes
Hum a phrase and the returned melody is often close but wrong: an interval off by a semitone, the ending turning the wrong way. The trouble is not pitch accuracy. Humming passes a contour, not notes. A text description passes a mood and not even a contour. MIDI passes pitch, duration and velocity, which gives the most control and also invites the model to smooth everything out.
Why MIDI conditioning holds better
Symbolic input breaks the melody into discrete events with explicit pitch and length, so the model does not have to guess from a spectrum. The cost is regularisation: pushed syncopations return to the grid, ornaments merge into the main note, a decaying long note becomes a fixed velocity. With MIDI, keeping the rhythmic spelling matters more than keeping the timbre; confirm the note count and bar lines first, then listen for expression.
Comparing the methods
| Input | What it locks | Usual failure | Check |
|---|---|---|---|
| Hummed, single line, no accompaniment | Contour, phrase length, accent placement | Semitones flattened, attacks vague | Play each interval on a piano |
| A single MIDI line | Pitch, duration, register | Syncopation straightened, ornaments lost | Count notes per bar, mark changes |
| A finished excerpt as reference | Timbre, character, arrangement density | The tune itself is borrowed, rights unclear | Take two bars, change three things |
| Described direction | Mood, rising or falling | Random landing, runaway range | Name the first note, the top note, the cadence |
An audio reference is the most misused input: it locks what something sounds like, not which notes are played. To keep your own melody, do not feed a whole chorus. To borrow a character, take two bars, change the ending and the rhythm, then describe the direction in words.
Limits of humming and audio references
Humming is good for starting and poor for signing off. Pitch drift, breath and slides all push the model toward a reading that is not in the score. An audio reference also drags in the arrangement texture and the reverb, so what comes back is a new song that resembles that one rather than your melody. Once the melody identifies the track, check it against the score before delivery: which bar the theme first appears in, and whether the two choruses really state the same tune.
Listen in the library
All tracks from midicn-lib — inline badge is the license tier (C1 commercial / C2 non-commercial / C3 study only).
Along the same line
Sources
- William Russo, Composing Music（Prentice-Hall）与 Hindemith, The Craft of Musical Composition 中关于旋律轮廓与句读的章节。
- 音频与符号两类旋律条件的差异，依据公开模型文档。