Speech source separation via independent low-rank matrix analysis with cepstrum-based complexity reduction

Independent low-rank matrix analysis (ILRMA) is a powerful blind source separation (BSS) method combining spatial modeling with a low-rank source model based on nonnegative matrix factorization (NMF). However, ILRMA often suffers from sensitivity to random initialization, which causes the block permutation problem and degrades separation performance in speech source separation. In this paper, we build upon CBD-ILRMA, an extension of ILRMA that incorporates cepstrum-basis-decomposition into the NMF-based source model to reduce the effective complexity of speech signals, thereby suppressing convergence to local optima and achieving more stable source separation. Separation experiments using several speech mixtures demonstrate that CBD-ILRMA consistently outperforms conventional ILRMA in both average separation performance and stability. In particular, CBD-ILRMA achieves mean SDR improvements of up to 6 dB over ILRMA, averaged over 100 independent trials. Furthermore, we investigate extended versions of CBD-ILRMA incorporating detail recovery strategies, which adaptively control the application of CBD during optimization. Experimental results indicate that these extensions can further improve separation performance. However, their effectiveness depends on the characteristics of the target sources, suggesting that appropriate adaptive control strategies are required. These findings highlight the importance of complexity control in robust speech source separation.