Dataset EDA

What the EDA panel shows

What is exploratory data analysis (EDA)?

EDA is looking at a dataset before modelling it: how balanced its labels are, how its acoustic features are distributed, and which clips are outliers. Skew or oddities in the data are easy to mistake for model bugs later, so this panel shows them first.

Acoustic features

Features are extracted per clip with librosa, and each is summarised as mean, standard deviation, min, and max:

• Spectral centroid, rolloff (85%), and bandwidth — where the energy sits in frequency and how wide it spreads.
• MFCCs — the first 13 coefficients.
• Chroma and tonnetz — pitch-class and tonal content.
• RMS energy and zero-crossing rate — loudness and noisiness.
• Tempo.

Clustering and nearest neighbours

The EDA panel also runs HDBSCAN clustering over the dataset's embedding space (default min_cluster_size 5, adjustable 2–50; at least 3 clips are needed) and reports a silhouette score for the resulting clusters (−1 to 1, higher is better separated), so you can judge how well-separated the clusters actually are rather than taking cluster assignments at face value. Clips HDBSCAN can't place in any cluster are labelled noise (-1) and are left out of the silhouette calculation. From any point, you can pull up its nearest neighbours in that same embedding space — useful for spotting near-duplicates or checking whether a mislabeled clip has neighbours that suggest what its label should have been.

This reuses the same embedding space as the Embedding projector panel — the two are looking at the same underlying representation from different angles.

Exporting results

There is no SRT, PDF, or transcript export anywhere in AudioLens — the exports above are the complete list.

Dataset-level EDA data is served by GET /{dataset}/eda for built-in datasets; see /docs/rest for the full endpoint reference.

Where to go next