🌐 3D Speaker Embedding Clusters: von Mises-Fisher Spread, Few-Shot Centroid Error and Bayes Overlap
Speaker embeddings as von Mises-Fisher clouds on a unit sphere: concentration, few-shot centroid arrows, nearest-centroid classification and the Bayes overlap of neighbouring clusters, with closed forms for resultant length and centroid error checked by Monte Carlo and quadrature (synthetic embeddings).
⚙ Under the hood
Speaker embeddings as von Mises-Fisher clouds on a unit sphere: concentration, few-shot centroid arrows, nearest-centroid classification and the Bayes overlap of neighbouring clusters, with closed forms for resultant length and centroid error checked by Monte Carlo and quadrature (synthetic embeddings).
Directional Statisticsvon Mises-FisherSpeaker EmbeddingCentroid EstimationBayes ErrorMean Resultant LengthUnit SphereClassification
3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install