用於語音增強之偽影感知加權損失函數

The Speech Enhancement (SE) system not only enhances the perceptual quality of speech but also make the ASR performance robust in noisy enviornments when integrating with ASR systems. However, single-channel SE may generate detrimental artifacts to ASR recognition, leading to recognition errors. Recent research indicates that by introducing the novel SE loss function NAaLoss and fine-tuning the model, the generation of artifacts can be effectively reduced. Nonetheless, this approach still needs to be revised in its underlying assumptions. Therefore, we extensively analyze this method in this study and conduct numerous experiments and case studies to identify the inconsistencies. To address this, we propose an improved loss function, AaWLoss. AaWLoss successfully resolves the potential loss of noise-condition artifact suppression inherent in NAaLoss under the same settings through modifications and optimizations. Furthermore, AaWLoss achieves peak performance in suppressing artifacts under clean conditions, even adding information beneficial for ASR recognition to the enhanced clean speech.