Convolutive NTF for Ambisonic Source Separation Under Reverberant Conditions

This paper presents a Non-negative Tensor Factorization (NTF) based sound source separation method with a novel convolutive Spatial Covariance Matrix (SCM) model, that is suitable for use with reverberant Ambisonic signals. The presented solution builds upon a previous work on SHD SCM-based NTF, but unlike the original, non-convolutive approach, it avoids the problem encountered when the analysis window is too short to capture the dominant part of the reverberant signal. Here we introduce a novel convolutive SCM model that accounts for reverberation which spans over multiple time frames and then we derive the corresponding parameter update equations. In particular, this work considers several variants of these updates, describes the underlying motivation for each algorithm design choice and indicates the update rules, which offer the highest gain in Signal-to-Distortion Ratio (SDR). The proposed solution is evaluated against the original approach for various reverberation time values, number of sources and types of source signals, using simulated first-order Ambisonic recordings. The results of this preliminary study clearly indicate that the proposed method enables higher quality of separation compared with the reference, non-convolutive algorithm.