Loading page
Visualize learned representations in a reduced space, with clustering and nearest neighbours.
The embedding projector reduces a model's learned representations down to a space you can actually look at — two or three dimensions you can rotate, zoom, and select points in.
An embedding is the vector of numbers a model uses internally to represent a clip. Clips the model treats as similar end up with similar vectors. Because these vectors are far too long to read, the projector squeezes them into 2D or 3D so you can see clusters, outliers, and clips that look mislabeled.
Each clip becomes one point. Its vector is the model's mean-pooled encoder output — Whisper's last encoder hidden state, or Wav2Vec2's 768-dimensional representation. The projection only reduces those vectors for display, so switching method re-runs the reduction without re-extracting embeddings.
The embedding operation supports three reduction methods:
Each can be projected into 2D or 3D, your choice.
Clustering is optional and off by default. When enabled, it runs
HDBSCAN over the projected points, with min_cluster_size tunable from
2 to 50. Smaller values surface more, smaller clusters; larger values
collapse the projection into a handful of broad groups (default 5).
Points HDBSCAN can't assign to any cluster are marked as noise, and the
result includes a silhouette score (−1 to 1) for cluster quality. At
least 3 clips are required.
Click a point and the panel can retrieve its nearest neighbours in embedding space — useful for answering "what else does the model think is similar to this clip?" without leaving the projector.