Speech extraction under extremely low SNR conditions

The acquisition of the target signal in noisy environments remains a prominent focus in the realm of signal processing. Although extensively studied, speech extraction under extremely low signal-to-noise ratio (SNR) conditions remains a formidable challenge. In this paper, we address this challenging task in the framework of independent vector extraction with the orthogonal constraint (OGIVE). We use real speech data to analyze the behaviors of the cost function under different SNR conditions, which inspires us to select appropriate parameters for optimization. Furthermore, we propose natural gradient-based algorithms to improve conventional OGIVE algorithms. Numerical experiments across various scenarios demonstrate the effectiveness and robustness of our proposed algorithms. •Speech extraction with extremely low signal-to-noise ratio (SNR) is rarely explored.•Theoretical analyses based on real data under different SNR conditions are conducted.•Algorithms optimizing the mixing vectors are advantageous under low SNR conditions.•Natural gradient algorithms for the demixing/mixing vectors are proposed.•The proposed algorithm demonstrates performance improvement and robustness.