Multiple DOA estimation based on estimation consistency and spherical harmonic multiple signal classification
Sina Hafezi, Alastair H. Moore, Patrick A. Naylor · 2017
A common approach to multiple Direction-of-Arrival (DOA) estimation of speech sources is to identify Time-Frequency (TF) bins with dominant Single Source (SS) and apply DOA estimation such as Multiple Signal Classification (MUSIC) only on those TF bins. In the state-of-the-art Direct Path Dominance (DPD)-MUSIC, the covariance matrix, used as the input to MUSIC, is calculated using only the TF bins over a local TF region where only a SS is dominant. In this work, we propose an alternative approach to MUSIC in which all the SS-dominant TF bins for each speaker across TF domain are globally used to improve the quality of covariance matrix for MUSIC. Our recently proposed Multi-Source Estimation Consistency (MSEC) technique, which exploits the consistency of initial DOA estimates within a time frame based on adaptive clustering, is used to estimate the SS-dominant TF bins for each speaker. The simulation using spherical microphone array shows that our proposed MSEC-MUSIC significantly outperforms the state-of-the-art DPD-MUSIC with less than 6.5° mean estimation error and strong robustness to widely varying source separation for up to 5 sources in the presence of realistic reverberation and sensor noise.