Federated Learning for Audio Semantic Communication
Federated Learning for Audio Semantic Communication
复制标题
DOI：
10.3389/frcmn.2021.734402
复制
发表时间：
2021-09
期刊：
影响因子：
--
通讯作者：
Haonan Tong;Zhaohui Yang;Sihua Wang;Ye Hu;Omid Semiari;W. Saad;Changchuan Yin
中科院分区：
文献类型：
--
作者：
Haonan Tong;Zhaohui Yang;Sihua Wang;Ye Hu;Omid Semiari;W. Saad;Changchuan Yin
本文研究了无线网络中音频语义通信的问题。在所考虑的模型中，无线边缘设备使用语义通信技术将大尺寸的音频数据传输到服务器。该技术允许设备仅传输捕获音频信号的上下文特征的音频语义信息。为了从音频信号中提取语义信息，提出了一种基于卷积神经网络（CNN）的波矢量（wav2vec）结构的自动编码器。所提出的自动编码器能够以少量的数据实现高精度的音频传输。为了进一步提高语义信息提取的准确性，在多个设备和服务器上实现了联邦学习（FL）。仿真结果表明，该算法能有效地收敛，与传统编码方案相比，可将音频传输的均方误差（MSE）降低近100倍。
In this paper, the problem of audio semantic communication over wireless networks is investigated. In the considered model, wireless edge devices transmit large-sized audio data to a server using semantic communication techniques. The techniques allow devices to only transmit audio semantic information that captures the contextual features of audio signals. To extract the semantic information from audio signals, a wave to vector (wav2vec) architecture based autoencoder is proposed, which consists of convolutional neural networks (CNNs). The proposed autoencoder enables high-accuracy audio transmission with small amounts of data. To further improve the accuracy of semantic information extraction, federated learning (FL) is implemented over multiple devices and a server. Simulation results show that the proposed algorithm can converge effectively and can reduce the mean squared error (MSE) of audio transmission by nearly 100 times, compared to a traditional coding scheme.