Multichannel non‐negative matrix factorisation based on alternating least squares for audio source separation system

Abstract

A multichannel blind source separation algorithm based on the multichannel non-negative matrix factorisation (NMF) model and an alternating least squares (ALS) method is developed. To develop the proposed algorithm, the multichannel NMF (MC-NMF) model is modified with stacked matrix notation. In the model, all parameters – frequency basis, time basis and mixing matrix – are estimated using the ALS method. The proposed MC-NMF algorithm is evaluated using an ‘underdetermined speech and music mixture’ dataset from the International Signal Separation Evaluation Campaign 2013 (SiSEC 2013). Experimental results show that the proposed algorithm outperforms the conventional NMF algorithms.