Underdetermined Blind Source Separation of Bioacoustic Signals

Bioacoustic signals have been used as a modality in environmental monitoring and biodiversity research. These signals also carry species or individual information, thus allowing the recognition of species and individuals based on vocals. Nevertheless, vocal communication in a crowded social environment is a challenging problem for automated bioacoustic recogniser systems due to interference problems in concurrent signals from multiple individuals. The bioacoustics sources are separated from the mixtures of multiple individual signals using a technique known as Blind source separation (BSS) to address the abovementioned issue. In this work, we explored the BSS of an underdetermined mixture based on a two-stage sparse component analysis (SCA) approach that consisted of (1) mixing matrix estimation and (2) source estimation. The key point of our procedure was to investigate the algorithm's robustness to noise and the effect of increasing the number of sources. Using the two-stage SCA technique, the performances of the estimated mixing matrix and the estimated source were evaluated and discussed at various signal-to-noise ratios (SNRs). The use of different sources is also validated. Given its robustness, the SCA algorithm presented a stable and reliable performance in a noisy environment with small error changes when the noise level was increased.