Spectral Clustering
A clustering method that constructs a similarity graph and uses eigenvectors of a graph-derived matrix to reveal non-convex group structure.
Machine-Learning Context
Spectral clustering builds a similarity graph and uses eigenvectors derived from its graph Laplacian to expose cluster structure that need not be convex in the original feature space. In speaker diarization it can operate on an embedding affinity matrix, where normalization, eigengap selection, and cluster-count estimation affect quality.
Terminology Boundary
The word “spectral” here refers to the eigen-spectrum of a graph-derived matrix, not to Fourier or STFT analysis of an audio spectrum.
Related Machine-Learning Concepts
- Diarization Clustering
- Speaker Embedding
- Cosine Similarity
- Graph Laplacian
Direct source: The primary paper or official specification for Spectral Clustering is linked here for verification.