Single-channel blind separation using L₁-sparse complex non-negative matrix factorization for acoustic signals.

Area of Science:

• Signal Processing
• Machine Learning
• Acoustics

Background:

• Single-channel blind source separation (BSS) is challenging due to signal ambiguity.
• Existing methods often struggle to preserve crucial signal characteristics like phase information.

Purpose of the Study:

• To propose an innovative single-channel blind source separation method.
• To enhance the factorization of source signals by preserving phase and optimizing sparsity.

Main Methods:

• A complex-valued non-negative matrix factorization (NMF) approach is utilized.
• Probabilistically optimal L1-norm sparsity is incorporated to enforce structure in temporal codes.
• An efficient algorithm with a closed-form expression for parameter computation, including sparsity, was developed.

Main Results:

• The proposed method effectively preserves phase information of source signals.
• Optimal sparsity enforcement leads to more meaningful parts-based factorization.
• Experimental validation using real-time acoustic mixtures demonstrates the method's effectiveness.

Conclusions:

• The developed complex-valued NMF with L1-norm sparsity offers an effective solution for single-channel BSS.
• The method provides improved source signal separation by leveraging phase information and sparsity.
• The efficient algorithm facilitates practical application in real-time scenarios.