Bone-Conducted Speech Codec Based on AMR-WB Framework and MHSA-CycleGAN Network
Xiaoqiang Hu, Zhe Chen, Fuliang Yin · IEEE Transactions on Audio Speech and Language Processing · 2024
The integration of Bone-conducted Microphone (BCM) speech codec and speech bandwidth expansion in a tandem manner often suffers from the complex structure and high computational complexity. To address this problem, a bone-conducted speech codec based on the AMR-WB (Adaptive Multi-Rate Wideband) framework and MHSA-CycleGAN (Multi-Head Self-Attention CycleGAN) network is proposed in this paper. Specifically, a new AMR-WB encoder structure with bandwidth extension capabilities for BCM speech is proposed. Then, the ISF (Immittance Spectral Frequencies) parameters from BCM speech are converted into Air-conducted Microphone (ACM) speech by deep neural network for extending the high-frequency components of BCM speech. Finally, a novel CycleGAN network with multiple attention mechanisms and loss functions incorporating high-order statistical characteristics is presented to capture the nonlinear relationships among ISF parameters. The proposed method can improve speech quality during BCM speech encoding and has lower computational complexity. Simulation experiments confirm that the proposed method consistently delivers effective speech enhancement performance across various bitrates for BCM speech and reduces at least 68% run time compared with state-of-the-art methods.