In a MIDI transcription of Burgmüller’s L’Arabesque, a research prototype produced a version that kept two separate sets of changes: eight accompaniment notes were removed, and twenty melody notes were shortened. Each set had first been applied to the original on its own. The combined version retained both requests, with the remaining events checked against the original.
I am exploring this around Battuta, the command-line MIDI editor I build. The question is what happens after making several versions of the same passage. If one experiment changes the accompaniment and another changes the melody, can their requested changes be carried into one file that is ready to compare with both?
The experiment answers a bounded part of that question. The inputs include the original file, both edited versions, and the exact requests that produced them. The combination remains a research prototype outside Battuta’s released command interface.
Four versions of the same passage
The source is a Mutopia transcription, and the selected range is bars 3–6 as represented in that MIDI file. A MIDI note has a start and a written duration; shortening that duration moves its release earlier without moving the start. The two experiments use that duration and the number of notes in the accompaniment as separate things to try.
Four versions, one passage
Which changes would you keep together?
Play the passage, then switch either change on or off. Each selection starts at the same listening position.
AOriginal
The accompaniment chords and melody keep their original notes and written durations.
4.74 seconds, including context around file bars 3–6. Switch while paused to keep your place, or while playing to compare.
All four versions and their files
A · Original
The accompaniment chords and melody keep their original notes and written durations.
B · Thinner accompaniment
One middle note is removed from each of eight accompaniment chords. The melody keeps its original durations.
C · Shorter melody notes
Twenty melody notes keep their starting times and end at 75% of their original written durations. The accompaniment is unchanged.
D · Both changes
The eight accompaniment deletions and twenty shorter melody notes are kept together.
B and C were each made from A. Their requests identify the original notes, so the prototype can take both sets back to that common starting point. D is constructed by applying the retained requests there.
Those requests give the work continuity across files. The accompaniment experiment remains a particular set of deletions; the melody experiment remains a particular set of shorter notes. Combining them does not require asking an agent to recreate either version from a description of how it ought to feel.
What the combined file actually contains
Before constructing D, the prototype replays each request against A and compares the resulting events with the supplied B or C. That establishes that the recorded request accounts for the version being offered. It then constructs the combined candidate in both request orders and compares their event streams.
For this passage, both orders produced the same result. A separate preservation check confirmed all twenty-eight requested edits and the other carried events. The later follow-up reconstructed the case five times, with the outputs agreeing with the earlier candidate and the inputs unchanged.
The useful result is a fourth file with a known relationship to the other three. Comparing A with B isolates the accompaniment deletion. Comparing A with C isolates the duration change. Comparing B with D puts that duration change into the thinner accompaniment. Those comparisons make different questions available for listening.
The existing core could execute these edits directly. The research adds the work around execution: checking the supplied versions against their requests, coordinating those requests, and examining the complete result. This case did not require a new language for musical intent.
Preparing the comparison for ears
For this page, all four complete MIDI files were freshly rendered through the same FluidSynth setup and soundfont before excerpts were taken. Rendering the complete files first retains the synthesiser’s earlier context when the selected passage begins. Each excerpt contains the changed bars with surrounding material. Switching versions keeps the same position in that passage.
The experiment has established that D contains the requested changes. No human listening decision is recorded here. Whether the thinner accompaniment and shorter melody belong together is a musical question the four recordings let a listener examine.
Two changes can interact even when they address different events and can be applied in either order. Separate research cases show why: extending a note can carry its release into a pedal interval changed by another request. Preserving each request does not preserve every relationship the music had before them.
That is a reason to put the combined version into the comparison. A, B and C establish the individual alternatives; D lets the listener judge their encounter. The files are ready for that decision without requiring either earlier experiment to be reconstructed.
Research record and listening scope
This is a research prototype using Battuta at source revision 7b5a0a8, outside its released command interface. The real-file and audio record describes the four versions and preparation. The mechanical follow-up records preservation checks and repeated construction; the research index connects the experiments.
The source is J. F. F. Burgmüller’s L’Arabesque, transcribed by Bas Wassink for Mutopia. The selected bars are those represented by the MIDI file. B removes eight notes; C shortens twenty notes; D retains both requests.
The page’s recordings are new renders of the four MIDI files checked against the research input hashes. They use FluidSynth 2.6.0 and GeneralUser GS, 44.1 kHz stereo, gain 0.2, chorus and reverb on, and no loudness normalisation. The same excerpt is cut from each full render with 20 ms fades at its outer edges. The rendering record gives file and soundfont hashes; the reproduction files include the script, MIDI inputs and instructions. The caller supplies the soundfont. No human listening decision or adopted combination policy is claimed.