CCLCap-AE-AVSS: Cycle consistency loss based capsule autoencoders for audio–visual speech synthesis

Abstract

Audio–visual speech synthesis (AVSS) is a rapidly growing field in the paradigm of audio–visual learning, involving the conversion of one person’s speech into the audio–visual stream of another while preserving the speech content. AVSS comprises two primary components: voice conversion (VC), which alters the vocal characteristics from the source speaker to the target speaker, followed by audio–visual synthesis, which creates the audio–visual presentation of the converted VC output for the target speaker. Despite the progress in deep learning (DL) technologies, DL models in AVSS have received limited attention in existing literature. Therefore, this article presents a novel approach for AVSS utilizing capsule network (Caps-Net)-based autoencoders, with the incorporation of cycle consistency loss. Caps-Net addresses translation invariance issues in convolutional neural network approaches for effective feature capture. Additionally, the inclusion of cycle consistency loss ensures the retention of content information from the source speaker. The proposed approach is referred to as cycle consistency loss-based capsule autoencoders for audio–visual speech synthesis (CCLCap-AE-AVSS). The proposed CCLCap-AE-AVSS is trained and tested using VoxCeleb2 and LRS3-TED datasets. The subjective and objective assessments of the generated samples demonstrate the superior performance of the proposed work compared to the current state-of-the-art models.

Keywords

• voice conversion
• audio–visual synthesis
• autoencoder
• capsule network
• cycle consistency loss