CATAD: Conformer-Based Adversarial Training with Adaptive Diffusion for Bone-Conducted Speech Enhancement

Due to the low-pass filtering of human tissues, the high-frequency components of bone-conducted (BC) speech suffer severe attenuation, resulting in reduced speech quality and intelligibility. To address this issue, this paper proposes a novel approach called CATAD, which employs conformer-based adversarial training with adaptive diffusion for bone-conducted speech enhancement. In CATAD, we deploy a TS-Conformer consisting of two Conformer modules that capture temporal and frequency dependencies, respectively, to enhance the expressiveness of speech features. To ensure the stability of the adversarial learning process and the diversity of the generated results, we adopt an adaptive diffusion process where random noise is added to the generated and real data, respectively. The proposed adaptive diffusion process boosts the capability of the discriminator to distinguish between real air-conducted speech and enhanced BC speech. Experimental results on the ESMB dataset demonstrate that the proposed CATAD can recover absent high components from BC speech, substantially improving speech quaility and intelligiblity of BC speech.