· 4 min read

Inside the DAFx Research Atlas

AIRAGembeddingsaudioDAFx

In my last post, I described how I collected the DAFx archive, made its papers searchable, and labelled them. This post is about the questions I could start asking once that work was done, and how far I think the answers can be trusted.

Questions the archive can already help me ask

I wanted to know which effects had attracted the most attention. That needed a rule: an effect had to appear in a paper’s title, or in its abstract and at least four stored text chunks. In this exploratory count, reverberation has 108 matches and distortion, overdrive or fuzz has 88. The groups overlap, and the matches have not all been reviewed. It is a place to start reading, not a final ranking. The cards below show the counts and the other questions I tried.

Seven ways into the archive. Each chart counts papers under a stated rule; a mention, a model proposal and a confirmed relationship are different kinds of evidence.

01Exploratory count

Which effect received the most research attention?

Reverberation leads this preliminary dedicated-paper search.

Papers matched by effect family
Reverberation108
Distortion / overdrive / fuzz88
Binaural / HRTF67
Source separation48

Rule: effect in the title, or in the abstract and at least four stored text chunks. Families overlap. This is not a final measure of research investment.

02Searchable mention

When did JUCE first appear?

2009earliest searchable mention

A click-synthesis paper names the JUCE C++ library in its implementation passage. JUCE appears in 49 readable papers.

First observed in this searchable snapshot, not proven first use anywhere.

03Use unverified

When did researchers start using librosa?

2015earliest searchable mention

The earliest match is an evaluation of audio-feature toolboxes. Librosa appears in 25 readable papers.

A mention can be a comparison or citation. I have not yet verified the first use in a paper’s own method.

04Tracked mentions

Which programming languages and frameworks appear most often?

C++ matters here: its exact name occurs in 207 readable papers, second among these six tracked names.

Distinct papers mentioning each name
MATLAB351
C++207
Python131
PyTorch82
JUCE49
TensorFlow / Keras30

C++ uses an exact-token match, including C/C++ and C++11. Counts cover the archive title, abstract or searchable non-bibliography text, not audited implementation choices. Names can co-occur, and the list is not exhaustive.

06Exact-name count

Which authors have the longest DAFx history?

Papers under recorded names
Vesa Välimäki71
Stefan Bilbao49

Recorded spans: 1998–2026 and 2003–2025, respectively.

Name variants are not disambiguated, and paper count is not a measure of influence.

07Open answer

Which DAFx paper is the most cited?

No defensible all-time winner yet.

A 2019 real-time black-box-modelling paper has 42 visually confirmed incoming links from DAFx papers in the identified subset.

Many bibliographies and candidate links still need review. The graph is useful for following documented connections, not for an all-time leaderboard.

What can I actually trust?

The questions above are short. Making their answers trustworthy was not. I could ask a model to label all 1,706 entries, but a paper that mentions a neural network is not necessarily a neural-audio paper. Sometimes the abstract leaves out the method. Sometimes the PDF text misses the passage that would settle it. Several models can agree and still be reading the same incomplete evidence.

That is why this was not a weekend project. The time went into deciding what each count meant, checking troublesome papers against the actual pages, and leaving uncertainty visible when the archive could not support a clean answer. I started out wanting a map of DAFx research. The more interesting result is a map I can question: one that shows me the papers behind a pattern, and where I still need to read for myself.