Non-negative matrix factorization based single-channel source-separation of passive underwater acoustic signals in deep sea.

Abstract

We use non-negative matrix factorization for source separation on ultra-low frequency passive-acoustic data from a single-channel recording acquired in deep sea. Non-negative matrix factorization decomposes the spectrogram into a spectral-component matrix and a time-encoding matrix. Detectors use known time-frequency features to group components from the same source and reconstruct spectrograms of blue whale calls, seismic sounds, and ship noise. Data are separated at low computational cost and without learning step. The separation assessment using scale-invariant signal-to-distortion ratio on spectrograms of simulated reference data is satisfying. Source separation on ocean-bottom seismometer data from the Southern Indian Ocean provides convincing results.