Too heavy, and I could not say which notes

I was listening to a MIDI rendering of Burgmüller’s Arabesque, Op. 100 No. 2. Something in the left hand bothered me, first in the opening and again about fifteen seconds in. This is what I was hearing:

I did not know yet that every one of those chords was struck at the same velocity. I only knew how it sounded. The opening felt heavy: the piano struck at even intervals with the same force each time, and the middle did the same. What I actually said was less precise than the feeling. It was too repetitive, too cluttered.

No bar numbers, no note names, no parameters. A feeling about sound, in ordinary language. I had not looked at the score.

I was working with an agent through OliviaMIDI, a tool that lets a person and a language model collaborate around the same MIDI file. The Rust core, Battuta, reads, modifies, and diffs MIDI through a command-line program called mid. By the end I had learned something about the piece I did not expect.

The opening is easy to locate. The similar part in the middle is not. It could mean an entire passage or only the accompaniment within it. Adjusting everything in range would drag the melody down with the chords.

The agent read the file and found the candidates: left-hand chords in bars 1–6, and a similar figure returning around bar 20.

Candidates, not targets. I needed to hear them.

The agent rendered two short trial listens. The second started one bar early, so I could hear where I was. A listen can be wider than the edit: that bar would be heard, not touched.

I listened, and could say more. Yes, that was it. I wanted the dynamics to curve: the opening should feel like an entrance, and the middle could stay even but alternate heavy and light.

The direction came from language. The object came from listening. The precise file locations were the tool’s job.

In the original file every one of those chords had velocity 69, the force of the key strike. The agent proposed a shape for each passage:

These numbers are one audible interpretation of my words. I set the direction; the agent proposed the values; my ear judged. The alternation’s heavy beats are also lighter than the original 69. I had not asked for that; it was the agent’s choice, and I kept it after listening.

Each passage has twelve chords of three notes: seventy-two notes modified in total. Pitch, timing, duration, and melody were preserved. The next listen could answer a clean question: with only these accompaniment velocities changed, how does it sound?

mid apply wrote a new MIDI file and kept the original. mid diff reported seventy-two velocity changes and nothing else, and a separate parser agreed. Every render used the same playback settings, cut from one full-length render. Whatever I heard differently could only come from those seventy-two notes.

I listened again. It was right.

What the score had to say

Afterward I asked a question I should have asked earlier: why did the composer write it that way? Were those uniform chords serving a purpose I was overriding?

The MIDI came from Mutopia, generated by LilyPond, and Mutopia publishes the printed score alongside it. The page marks the opening piano, puts staccato dots on the first two bars of chords, and writes a crescendo. The MIDI rendered all of it at a flat velocity of 69, with every chord sustained to the next beat.

Bars 1–6, twice

What the page asks for, what the file played

Printed score

p at bar 1staccato dots, bars 1–2cresc. at bar 5

MIDI file

69

Twelve chords at velocity 69, each sustained to the next beat.

What I heard as too heavy, too repetitive, was not entirely a preference. Part of it was information lost when notation became a MIDI file. The staccato and dynamic shaping existed in the score. They were not in the rendering.

That changes what the edit was. Not only taste: in one parameter, it compensated for expression the rendering had discarded. The numbers were still an interpretation; they did not restore the staccato or follow the written crescendo. But the direction — lighter, graduated, varied — had been in the score all along.

I could have started by reading the score. Instead I started by listening, noticing something wrong, and saying so without technical vocabulary. From there: a trial listen to find where, numbers to make the direction audible, a tool to change exactly those notes, a check that nothing else moved, and another listen.

The score discovery came after. That order is the one I want. I would rather start from what I hear, change one thing, hear the result, and bring a more specific question to the notation, than try to implement the score faithfully without knowing what bothered me.

This was two passages, one parameter, one piece. It does not generalize to every vague description a listener might offer. The next ambiguous “that part” will still need a trial listen and a confirmation before anyone edits anything.