A finished Sound Cells volume contains several distinct output types: a PDF score in standard music notation, MP3 audio files of the exercises and compositions, MIDI files for playback and further processing, and a metadata manifest linking all the above to the mathematical structure (Forte numbers, interval vectors, subset indexes) of the trichord pair. Each of these outputs is produced by a different tool, connected by a Python orchestration layer that manages the full pipeline.

Understanding this pipeline is both a practical guide to music technology production and an example of how modern software engineering principles — modularity, separation of concerns, automation — apply to creative work at scale.

The Full Pipeline

The Role of MIDI

MIDI (Musical Instrument Digital Interface) is a 40-year-old protocol that encodes musical events — note on, note off, pitch, velocity, duration — as numerical messages. Despite its age, MIDI remains the lingua franca of music technology because it is universal, compact, and precisely defined.

In the Sound Cells pipeline, MIDI serves as the intermediate format between score generation (LilyPond) and audio rendering (SuperCollider). LilyPond generates a standard MIDI file as a byproduct of score compilation — every note in the score becomes a MIDI event with the correct pitch, duration, and timing. SuperCollider can read this MIDI file and use it as a sequencer to drive audio synthesis, ensuring that the audio matches the score exactly.

MIDI also serves as the format for the exercise files included with each Sound Cells volume — students can import them into any MIDI-capable software or hardware sequencer, play them back at any tempo, and transpose them to any key without affecting the pitch relationships. The mathematical precision of the pitch class set system is preserved exactly through the MIDI format.

LilyPond — Music as Code

LilyPond is unique among music notation programs in treating music as a compiled language: you write a text description of the score, and LilyPond compiles it to a PDF. This makes it ideal for automated generation — the same Python script that generates the pitch class arrays can also generate the LilyPond source code.

% LilyPond source — auto-generated by Python
% Trichord pair: 027-027, Key: C
version "2.24.0"
relative c' {
  clef treble
  key c major
  time 4/4
  % Trichord 1: 027 (C, D, G)
  c8 d g4 ~ g8 d c4
  % Trichord 2: 027 transposed (E, A, B)  
  e8 a b4 ~ b8 a e4
  % Combined hexatonic ascending
  c8 d e8 a b4 g2
}

The LilyPond source is human-readable — a musician can look at it and understand immediately what pitches are being specified — but it is also machine-writable. The Python template that generates the above source needs only to know the trichord pair and key center; the rest is boilerplate that stays constant across all 9,240 permutations.

File Format Relationships

What “Automation” Actually Means

Describing this pipeline as “automated” can create a misleading impression — that the music produces itself. In practice, automation handles the mechanical tasks (transposing pitch arrays, compiling LilyPond files, encoding audio) while human judgment governs every decision that affects musical quality.

The decisions that are not automated: which trichord pair to use for each volume, what exercise formats best illustrate that pair’s character, what tempo and instrument sound best serves the audio files, whether a specific generated exercise makes musical sense or reveals a problem with the specification. These require a musician’s ear and decades of experience with the material.

Automation provides scale — the ability to produce complete, consistent, high-quality materials for 78 volumes where manual production of each would be prohibitive. It does not provide musical judgment. The two work together, each doing what it does best.