Filtering as Cultural Technique

When examined from the perspective of signal processing, contemporary audio techniques such as style transfer and neural synthesis might best be understood not as hybrids of sampling and synthesis but as data-driven manifestations of digital filtering. From the smallest building block of neural networks to the source-filter operation that describes style transfer, filtering in synthetic media is operative as at once operation, concept, model, and metaphor. WP4 asks: how does attention to the cultural technique of filtering illuminate the relations between twentieth-century synthesis and twenty-first-century synthetic media?

The pliability and subsumptive quality of filtering is a constant within audio signal processing, predating synthetic media. That sound sources, room acoustics, the hearing function, sampling, and synthesis could all manifest instances of a single model raises the question of how meaningful a cultural technique filtering is in the first place. Yet the capacity of machine learning to move across domains and produce successful results, from text to image to audio, testifies to the domain-agnosticism and portability of contemporary signal processing, and thus perhaps to the power of meta-models such as filtering.

Approach

WP4 maps and analyses the growing hegemony filtering comes to assume in engineering practice against the changing horizon of industrial needs and commercial desires that twentieth-century signal processing responded to. A large corpus of published texts is assembled from research journals and canonical textbooks, analysed in historical periods: from the first observation of frequency-dependent responses in electrical networks (Heaviside, 1887), through the earliest articulation of digital filtering (Robinson, 1954), the consolidation of signal processing as a discipline in the 1970s, and the use of filters as the central operational block of perceptrons and neural networks.

Following WP2's NLP approach, topic modelling tracks which methods consolidated filtering as a master-concept, and which alternatives emerged and disappeared. Novel visualisation methods trace the movement of the filtering metaphor across domains, connecting current interest in translatability in machine learning systems to foundational concerns within signal processing.

Methods

Archival research across 15 archives including AT&T/Bell Laboratories, IEEE, Hagley Museum, and The Royal Society. Topic modelling (TOMOTOPY), natural language processing, optical character recognition (Tesseract), discourse analysis.

Outputs

Research articles; a finding aid for key archives; contributions to the Synthesis Handbook; a monograph situating filtering between synthesis, sampling, and industrial production.