Accurate Delayed Source Model for Multi-Frame Full-Rank Spatial Covariance Analysis

Multi-frame Full-rank Spatial Covariance Analysis (mfFCA) is a technique for a blind source separation method and can be applied to reverberant underdetermined conditions where the sources outnumber the microphones and the reverberation time is long. This model, however, does not express all direct and delayed source components in multi-frame observation vector. This paper proposes a new model that takes into account accurately the direct and delayed source components, by introducing delay-wise spatial covariance matrices. We have then derived new expectation-maximization and multiplicative update algorithms for the proposed model. Experimental results show that the proposed method performed better than the conventional mfFCA for the task to separate three sources with two microphones.