To meet the requirements of noise suppression with extremely low signal-to-noise ratios in multi-source coupled strong background noise scenarios, it is necessary to propose solutions combining learning-based technical means to compensate for the shortcomings of conventional algorithms in terms of adaptability. In the practical application of this technology, the following technical difficulties need to be solved: (1) The capacity limit of the current general network for noise suppression is −5 dB, and the lower boundary needs to be widened to meet the requirements of this task; (2) Due to the limitation of cost, the number of samples collected in the actual environment is very small, and there is no open source data available in similar scenes, which leads to poor training effect and over-fitting. To address these issues, this paper proposes an end-to-end supervised learning framework that involves dataset construction, network transfer, and training optimization. Test results demonstrate that the proposed solution is effective in suppressing coupled noise under extremely low SNR. Data augmentation methods for small samples further improve network performance.