A direct-path relative transfer function learning based sound source localization method is proposed to deal with the problem that localization features are easily distorted by noise and reverberation.

It embeds binaural signals to a real-valued direct-path feature space, which disentangles localization cues from other factors including noise, reverberation, etc.

The proposed method owns a strong robustness for sound source localization, and a good generalization ability for unseen binaural arrays.

A time-frequency-wise spatial spectrum clustering based multiple sound source counting and localization method is proposed to solve the problem that signals of multiple sources are mixed and the number of sources is unknown.

ŸWithout the priori of source number, it obtains the spatial spectrum clusters of multiple sources by iteratively detecting new sound source and adjusting the dominance association between sources and spatial spectra.

The proposed method achieves a superior performance for joint sound source counting and localization, and can localize sources with small angular separation.Ÿ   

A direct-path phase difference learning based multiple moving sound source localization method is proposed to deal with the problem that the position of moving sources is time-varying and sources are intermittently sounding.

It learns competing and time-varying direct-path phase differences for multiple moving sources，and avoids the problem of assignment ambiguity and uncertain output dimension encountered by the multi-target regression framework.

The proposed method is superior for the azimuth and elevation estimation of multiple moving sources, and the constructed spatial spectrum exhibits reliable peaks around the actual directions of sources.