A weakly supervised learning method based on optimal transport for sound sources reconstruction in the strong interference environment

Reconstruction of sound sources in the strong interference environment is difficult due to the low signal-to-noise ratio of muti-channel signals recorded by the microphone array. In this research, a loss function based on optimal transport is derived to suppress background interference for sound source reconstruction. The approximated distribution of uncontaminated data is obtained by training the snapshot matrix through weakly supervised learning, which means there is no need to collect paired data. The quantitative reconstruction of the actual amplitude of the sound source is realized by a normalization strategy with the inverse tangent function. In the numerical simulation and experiment, the proposed method is able to accurately reconstruct the sound pressure level of target sound sources in the strong interference environment.