Deep Clustering for Source Separation

Definition

Source separation approach where a neural network learns to embed each time-frequency bin into a space where bins belonging to the same source cluster together. Separation is performed by clustering the embeddings, then using cluster assignments as masks.

Key Ideas

• Hershey et al. (2016). Each T-F bin gets a D-dimensional embedding.
• Training objective: minimize distance between embeddings of bins belonging to the same source, maximize distance between different sources.
• At inference: run k-means on embeddings to get cluster assignments, use as binary masks.
• Elegantly handles the permutation problem — clustering is permutation-invariant.
• Foundation for later work: deep attractor networks, anchored deep clustering.

Relationships

• Introduced by john-hershey
• Related to ../concepts/permutation-invariant-training — alternative solution to permutation ambiguity
• Implemented in ../entities/nussl
• Preceded ../entities/demucs-style mask inference approaches

Sources

None ingested yet — seed batch setup.