Scribe: The composer as palaeographer

By Mark Dyer

In a recent blog, I introduced the first stages of Scribe, a creative project in which I trained a neural network (a form of machine learning) on pages of the medieval ‘Old Hall’ music manuscript.

Presented with the machine’s outputs – thousands of glitchy images with increasing likeness to the manuscript – I began working under the pretence the network was actually a medieval scribe, copying out ‘new’ compositions, each image a fragment of a fabricated facsimile manuscript .

I then acted as palaeographer – one who assembles and deciphers ancient texts – and started pairing and arranging images into the pages of this manuscript. As palaeographer, I needed to make sense of these strange images, following semi-regulated rules. Below is an example ‘page,’ compiled from four images - clockwise from top left, these are epochs 4010(1), 4005(1), 4005(0) and 4010(2).

‘Page’ of Scribe score, compiled of four machine-learning generated images, epochs 4010(1), 4005(1), 4005(0) and 4010(2)

Roughly, I aligned the red stave lines, left to right. The marks on the upper margins perhaps hint at the manuscript binding beneath, whilst the C-clef in the top-left corner of image 4010(2) (indicated in the second example below) suggests the left-hand side of the page.

Before transcribing the notation, I had to interpret my invented page. I identified three musical voices. The cantus is written faintly over one system with its own text at the bottom of 4010(1) and 4005(1):

The countertenor and tenor are paired across two systems of repetitive figures in 4010(2) and 4005(0), sharing a separate text.

I then transcribed individual voices, a balancing act between rigorous transcription following Willi Apel’s The Notation of Polyphonic Music 900-1600 (2010) and editorial license, jigsawing voice durations and harmonies together.

Let’s look at the first bars of the example page. Given the clef in the countertenor, I added clefs to the other voices to create an appropriate opening harmony. Next, I calculated the mensuration (equivalent to a modern time signature) of each voice: a back-and-forth process navigating the ambiguities of mensural notation and the low-resolution images.

For instance, interpreting a skewed notehead in the countertenor as a square breve would suggest one mensuration and ensuing harmony with the tenor, and as a diamond semibreve a separate mensuration. Each reading would open and close myriad possibilities. I see this iterative patching as a playful dialogue with the machine-generated notation, an affective agent in the creative process of Scribe.

In the next post, we’ll look at bringing this notation to life with EXAUDI vocal ensemble.