The Research on the Tibetan Speech Enhancement Method of MP-SENet Combined with CBAM and Multi-Scale STFT Loss

The phase information significantly affects speech quality and intelligibility. To address this, we optimize the MP-SENet model, which enhances both amplitude and phase spectra in parallel. The model extracts fused features through multiple paths, adjusts feature channel weights using the SE module, and incorporates CBAM attention and multiscale STFT loss. These improvements boost the model's ability to handle complex tasks. Experimental results show that the enhanced MP-SENet performs well in speech denoising, improving listening quality by explicitly estimating phase. Tests on the Tibetan and MUSAN noise datasets show a PESQ score of 3.214, outperforming existing phase-aware methods.