Build log · 2026.08.08

I gave a talk on audio ML in FiftyOne

Thirty minutes on curating, searching, and evaluating audio datasets — plus the two repos that came out of it.

fiftyoneaudio

On Thursday I spoke at the Voxel51 Audio and AI Meetup. My talk was “Curating, Searching, and Evaluating Audio Datasets in FiftyOne” — about thirty minutes on treating audio as a first-class citizen in a tool most people use for visual data.

There were three parts:

  • Similarity search over ESC-50 environmental sounds, using CLAP embeddings to query a collection by example clip or by natural-language prompt.
  • Captioning on Clotho: generating captions with CoNeTTE, then scoring them against the human references.
  • Moment retrieval, the Audio Moment Retrieval challenge from DCASE 2026 — finding a described moment inside a long recording and putting it on an interactive timeline.

The recording is on YouTube.

Two public repos came out of it. fiftyone-audio-toolkit is the plugin: a spectrogram renderer so clips are actually visible in the sample grid, an embeddings similarity-search panel, and the operators that compute the embeddings. fiftyone-audio-meetup has the slide deck and the dataset loaders/inference scripts, if you want to reproduce any of it.