Power Spectral Density Estimation for Acoustic Source Separation Using A Spherical Microphone Array

Acoustic source separation using microphone arrays has garnered increasing attention within the audio signal processing community. Cascading a beamformer and a post-filter for source separation is widely recognized as an effective solution, where accurate estimation of the power spectral density (PSD) is crucial for post-filter. However, estimating the PSD components is a challenging task in the presence of noise. In this paper, we propose an efficient multi-source PSD estimation method using a spherical microphone array (SMA) in noisy environments for source separation. Specifically, the orthogonality of spherical harmonics (SH) inherent in an error-free sampled SMA enables effective suppression of noise components in the spatial covariance matrix (SCM) of the observation signals, facilitating accurate PSD estimation for each individual source, thus the post-filter would be effectively constructed. The source separation problem is then addressed by applying a SH-domain beamformer along with the developed post-filter. Both simulation and experimental results demonstrate the proposed algorithm’s effectiveness over the baseline methods.