Search sound, not words.

Type a sound, get the sound

For each ready-made search: how many of the top 10 clips carry the right human label. Click a dot to run it.

The model hears the mood. Words can't reach it.

CREMA-D actors say the same sentences in six moods, so a transcript is identical. Typing the mood finds nothing; clicking a clip finds its kind. Bars: how many of an average clip's 10 nearest neighbours share its mood.

Three collections, one space

Every clip sits in the same space, whatever it came from. The layer buttons at the top hide any of them.

One model, two inputs, one space

Nothing is transcribed or tagged. Sound and words become the same kind of vector, and search is just “nearest”.

• Search by example skips words: it ranks clips by how close their vectors are to the clip you picked.
• ✓ / ✗ come from human labels the search never sees. Podcasts have no clip labels: ? means listen and judge.
• Embeddings come from two finished sister studies (audio-vocab, podcast-vibe-search). Nothing was trained for this page.

What the numbers don't show

• Labels are strict. A ✗ can still sound right: a giggle counts as laughter, but a crowd laughing inside an “applause” clip doesn't.
• Neighbours confuse neighbours. “Clearing throat” and “shouting” reach only 4 of 10: their nearest clips are coughs and screams. Here one plain word beat a richer description (0 → 4 of 10).
• The example chips are the best case. “Angry clip → similar” starts from the most typical angry clip; chart 02 shows the average.
• Podcast clips are 20 s long, the others 2–10 s, so a short cough inside a podcast is diluted. Podcasts also cluster by show (same hosts, mic and music).
• FSD50K audio plays once its zip is imported (it was embedded in Colab).