To work with prism-samplernn I began to build my concrete training dataset, a detailed and intensely focused sampling experience requiring a minimum of ninety minutes of material of individual sound objects for training. In order for the model to generalise and predict there is a process of statistical clustering and feature extraction, this helped me to decide on distinct classes of data:
- AGE OF ELECTRICITY_Recordings of noise, DIY electronics and electromagnetic frequencies
- ECHOES OF INDUSTRY_Recordings of Manchester mill spaces, tools and machinery
- MATERIALITY_Object recordings of metal sound sculptures and glass fragments
Throughout the process I was guided by Pierre Schaeffer’s ‘Notes on a Concrete Music’, reading again his experiments of sound object extraction and reflections on matter and form. In tandem I was reading ML textbooks on what comprises a ‘good’ dataset. These texts informed my workflow twofold, on the one hand trying to create a varied and distinct dataset but in other ways ‘challenging the machine’ and forcing errors. One example of this was with noise. My reading had told me to eliminate noise (in a statistical sense) wherever possible, also my trip to Russia had taught me that a neural synthesis model would be able to recognise attacks more than drones, events more than textures. As an artist who works with noise I wanted to test this and used noise purposefully within my dataset, picking particular machine timbres, pitches and electromagnetic textures which on reflection became the grounding of the sonic palette on output.
MATERIALITY DATASET EXAMPLE
Video: Glass fragment extraction, using a range of glass from National Glass Museum Sunderland
CONCRETE DATASET RULES
Throughout sound object extraction I worked to a set of self-directed rules for the dataset:
- ORIGINAL RECORDINGS: All recordings are of known material origin and self recorded
- EXTRACTION: Activate objects and spaces to uncover a wide range of resonances and frequencies with a variety of attacks, drones, durations and pitches.
- RECORDING: Contact mics, electromagnetic detectors, consideration of depth and space, different amplitude levels.
- NO MELODIC OR FORCED RHYTHM: Concentrate on the texture of the sonic fragments
- RAW MATERIAL: No effects added to the material or over production
- NOISE: a) reduction on MATERIALITY fragment recordings and b) purposeful inclusion of noise within ‘age of electricity’ machine data class
ECHOES OF INDUSTRY DATASET EXAMPLE
An ode to my hometown of Manchester, birthplace of the industrial revolution and a city defined by music, mills and radical politics, I wanted to capture the aura of these mill spaces, once used for textiles and industry – then occupied by music and art – and presently increasingly at risk from property developers and gentrification. One mill space that isn’t under threat and is a shining light for dynamic and subversive art is Salford’s Islington Mill, an ever evolving creative space, arts hub and community. The mill is currently at an exciting stage of its 200 year history with a transformative building renovation following huge fundraising efforts from the Mill community.
I was privileged to gain access to the legendary 5th floor space, days before this part of the building was closed and a new chapter began as building work started. Between this space, Wellington House Mill in Ancoats, and Kunstruct’s workspace at Rogue Studios, I recorded the sounds of machinery, tools, paper, slate, discarded pianos, ambient and electromagnetic recordings, contact mic recording of pipes and cavities, mechanism sounds of industrial hoists and most importantly spent time with the mill custodian pigeon dwellers. These recordings were imbued with a sense of place, emotion and industry, could a machine generate a sense of space and heritage?
One key observation in building the dataset was around labour, the perception that machine learning simply automates everything for you, I found to be misplaced, My experience of creating the dataset was so completely human, selective and labour intensive with the training data being the bedrock of the process and a space where the artist has complete control. I experienced a sensation of mirroring the machine, and through learning how the ML model trains was directly influencing my workflow and approach the building of the dataset methodically. I began to categorise and label, trying to predict or second guess the features it might detect and extract, whilst also disrupting this to force errors and unexpected outcomes in my choice of sound object.