Apppublic

eustlb/audio-classification-models-on-the-hub

sourceHugging Faceupdated 13h agoView on Hugging Face

0likes

App README

Number of audio classification models on the Hub

Every public model repo currently tagged audio-classification on the Hugging Face Hub, counted by the month it was created, with landmark releases pinned along the timeline as their lab's logo.

Hover a column for that month's figures, or click it to keep the box open. Hover a logo for the model; click it to open the repo on the Hub. Hovering a column gathers that month's logos underneath it.

Nothing runs on the Hub: this is a static Space, so no hardware is assigned and no container is started. The browser renders a page built ahead of time, and the figures sit in it as two <script type="application/json"> blocks.

Files

File	Role
`index.html`, `style.css`	the page, pre-rendered
`task.json`	which tag this Space charts, and its title
`data.json`	monthly counts: `[month, new, releases with 25+ likes]`
`releases.json`	the landmark releases and their lab logos
`fetch_data.py`	re-reads the counts from the public Hub API
`select_releases.py`	picks the landmarks and resolves their logos
`build_static.py`	writes `index.html` + `style.css`
`refresh.py`	does all three, then pushes what changed

Staying current

A single scheduled Hugging Face Job refreshes every audio-task Space in this family, this one included. It pulls each Space, re-reads the Hub API, rebuilds, and pushes back only files whose bytes changed — so a quiet day leaves no commit.

bash

python3 refresh.py     # or run it by hand

How landmarks are chosen

Ranked by likes, which is the only engagement signal the listing API returns, then filtered twice: repos that repackage another team's weights (GGUF, ONNX, MLX, quantisations) are dropped as formats rather than releases, and one owner may hold at most four slots so a single lab cannot crowd out the field. A pins.json, where present, is a hand-curated list merged in ahead of those.

Caveats

This is a survivorship count. The Hub API returns only models that exist today and carry the tag now. Repos deleted since, or created under another tag and retagged later, are invisible, so early months are understated.

The window starts at 2022-04 — it 2022-03 is the Hub's createdAt backfill, not activity: 75-100% of it lands on 2022-03-02 across hundreds of unrelated owners, so those repos predate the field. Cumulative figures in the hover box still count from the true start of the series.

Landmark figures are all-time, so older releases have had longer to accumulate downloads and likes than recent ones.