A Novel Loss Incorporating Residual Signal Information for Target Speaker Extraction Under Low-SNR Conditions

Most target speaker extraction (TSE) methods primarily focus on scenarios characterized by conventional signal-to-noise ratio (SNR), overlooking the substantial decline in the quality of extracted speech under low-SNR conditions. To address this issue, we propose to add an item related to the residual signal to the original SI-SDR-based loss function, with no need of additional training labels or network modules. Specifically, we investigate the impact of residual signal information by applying three types of constraints, which are direction, Euclidean distance, and projection distance + Euclidean distance. The proposed methods are experimentally evaluated using the WSJ0-2mix-extr datasets, with SNR range selected as -10 to -5 dB, -15 to -10 dB and -20 to -15 dB, and the SpEx series models are selected as the validation models. Experimental results show that the proposed methods are all superior to the baseline, and the performance improvement of the first two methods increases with the decrease of SNR, but the third method shows better performance when SNR is higher.