Advanced Blind Source Separation via Enhanced Sparse Adaptive Decomposition With Non-Negative Matrix Factorization: Benchmarking Performance in Audio-Visual Signal Decoupling

Blind Source Separation (BSS) plays a crucial role in signal processing, enabling the extraction of individual sources from mixed signals without prior knowledge of their origin. This capability is essential in applications such as speech enhancement, hearing aids, multimedia forensics, and human–computer interaction. Traditional approaches, however, often struggle with noisy environments, overlapping frequency components, and highly correlated audio-visual data streams. While Independent Component Analysis (ICA) and conventional matrix factorization methods have achieved noTable.success, their performance often degrades when signals exhibit sparsity or when temporal dependencies are nonlinear. In particular, mixed audio-visual data pose challenges due to the presence of redundant information, cross-domain interference, and the demand for high reconstruction accuracy. This study introduces an Enhanced Sparse Adaptive Decomposition (ESAD) framework integrated with Non-Negative Matrix Factorization (NMF) to address these limitations. The ESAD component adaptively enforces sparsity constraints, ensuring that the decomposed sources are well-separated and less prone to interference. NMF is then applied to extract meaningful latent structures, leveraging non-negativity to maintain physical interpretability of both audio and visual features. Together, the hybrid approach exploits both the sparsity and the structural coherence of the signals. Results showed a 15–20% improvement in separation accuracy and a noticeable enhancement in the intelligibility of speech under noisy conditions.