Attention-Enhanced Voice Portrait Model Using Generative Adversarial Network  

作　　者：Jingyi Mao Yuchen Zhou YifanWang Junyu Li Ziqing Liu Fanliang Bu 

机构地区：[1]School of Information Network Security,People’s Public Security University of China,Beijing,100038,China

出　　处：《Computers, Materials & Continua》2024年第4期837-855,共19页计算机、材料和连续体（英文）

基　　金：the Double First-Class Innovation Research Projectfor People’s Public Security University of China (No. 2023SYL08).

摘　　要：Voice portrait technology has explored and established the relationship between speakers’ voices and their facialfeatures, aiming to generate corresponding facial characteristics by providing the voice of an unknown speaker.Due to its powerful advantages in image generation, Generative Adversarial Networks (GANs) have now beenwidely applied across various fields. The existing Voice2Face methods for voice portraits are primarily based onGANs trained on voice-face paired datasets. However, voice portrait models solely constructed on GANs facelimitations in image generation quality and struggle to maintain facial similarity. Additionally, the training processis relatively unstable, thereby affecting the overall generative performance of the model. To overcome the abovechallenges,wepropose a novel deepGenerativeAdversarialNetworkmodel for audio-visual synthesis, namedAVPGAN(Attention-enhanced Voice Portrait Model using Generative Adversarial Network). This model is based ona convolutional attention mechanism and is capable of generating corresponding facial images from the voice ofan unknown speaker. Firstly, to address the issue of training instability, we integrate convolutional neural networkswith deep GANs. In the network architecture, we apply spectral normalization to constrain the variation of thediscriminator, preventing issues such as mode collapse. Secondly, to enhance the model’s ability to extract relevantfeatures between the two modalities, we propose a voice portrait model based on convolutional attention. Thismodel learns the mapping relationship between voice and facial features in a common space from both channeland spatial dimensions independently. Thirdly, to enhance the quality of generated faces, we have incorporated adegradation removal module and utilized pretrained facial GANs as facial priors to repair and enhance the clarityof the generated facial images. Experimental results demonstrate that our AVP-GAN achieved a cosine similarity of0.511, outperforming the performance of our compariso

关 键 词：Cross-modal generation GANs voice portrait technology face synthesis 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]