· 4 min read

Mapping 29 Years of Digital Audio Effects Research

AIRAGembeddingsaudioDAFx

In my last post, I built a study tutor around one 68-page PDF, Discover Canada. It was a useful first RAG project because I could check the model’s answers myself. When something went wrong, I usually knew why.

Then I picked a collection where that approach would not scale: the proceedings of DAFx, the International Conference on Digital Audio Effects. I use DAFx papers often and know a few people in the community. I wanted to know what the archive could tell me about the field as a whole. Which effects keep attracting new methods? When did learning-based approaches become central rather than incidental? Which older papers have surprising connections to recent ones? And who or what has had the most influence?

I decided to start by collecting the archive, making the papers searchable, and finding a way to explore nearly three decades of research: a kind of DAFx atlas. This post is about that process and the challenges I ran into along the way.

The DAFx archive

The official DAFx paper archive has 1,706 entries from 1998 through 2026 in the snapshot I used. Most are research papers; a few are talk abstracts. I could collect the titles and abstracts directly from the archive, without opening a PDF. A few entries have only a title.

From those I made embeddings: lists of numbers that represent what a piece of text is about. Two texts about similar problems end up with similar numbers, even when they use different words. That lets me search the archive by idea instead of by keyword.

Abstracts only go so far, though. I also wanted to search inside the papers for technical details, so I worked through the PDFs, splitting each one into small pieces of text that remember their page numbers. In the end, 1,685 of the 1,706 entries made it into that searchable index, at least in part.

How much of each year’s text could I actually use? The chart below shows every year’s papers, coloured by what I managed to extract: dark for text I could use, orange for text I could read but not fully trust, and grey for papers I couldn’t index. Hover over a year for the exact counts.

When PDFs weren’t machine-readable

Indexing the PDFs was harder than collecting the metadata. A paper can look fine on screen while its extracted text is broken. Two 2006 Csound papers, Cecilia and TclCsound and Streaming Frequency-Domain DAFx, turned some code into spaced-out letters. After checking the pages, I used paper-specific rules to join them. For a 2000 waveguide-mesh paper, I used OCR but indexed only the passages I could check against the page images.

That is what the orange and grey in the chart mean. Finding a paper is not the same as trusting every equation or number in its extracted text.

Giving the papers useful labels

Making the papers searchable wasn’t enough to understand how the research evolved. A paper might mention a neural network without actually using one. I needed to distinguish what each paper was about from how it approached the problem.

So I started assigning two labels: a research area (such as reverberation or virtual analog modeling) and a method family (such as classical DSP or deep learning). I tested the categories on 20 papers, revised the ones that overlapped or were unclear, and ended up with 14 research areas and nine method families.

I also wanted to know whether AI was central to a paper or only played a supporting role. That’s not always obvious from the title or abstract. For example, this 2025 paper uses differentiable DSP, although its title doesn’t mention it.

These classifications are still model-generated proposals. When something looks uncertain, I go back to the original paper rather than treating the label as a fact.

Where this led me

I started out wanting to make a large collection of papers searchable. As I worked through the archive, I became more interested in what I could learn from the collection as a whole.

Embeddings can suggest related papers, even across decades and vocabularies. But similarity doesn’t tell me whether one paper influenced another, whether a method was central to a paper or only mentioned, or whether a topic is really growing. For that, I still need the original papers and a closer reading.

So the project is gradually becoming an interactive atlas: somewhere I can explore research topics over time, find related work, and follow every result back to its source. In the next post, I share some of the first things I found.

I’m curious what I haven’t thought to ask yet. What would you want to know about 29 years of DAFx research? Email me with your suggestions.