Spectrally and spatially informed noise suppression using beamforming and convolutive NMF

Speech enhancement in low SNR conditions or in presence of large amount of reverberation is a challenging task. However, in some applications, prior information about the interfering noise source is available and can be exploited to tackle this issue. We propose to combine a beamformer with convolutive NMF in order to estimate the PSDs of the target speech signal and of the noise to be suppressed by exploiting knowledge of the noise source location and about its spectral content. We apply the proposed system to ego-noise suppression for a robotic platform. Simulations show that the spectral information exploited using convolutive NMF is beneficial to the noise reduction performance when compared to methods based on blind estimation but that estimating the noise PSD from the output of the beamformer is beneficial mostly when no prior knowledge of the noise spectral content is available.