Wav2Sem: Plug-and-Play Audio Semantic Decoupling for 3D Speech-Driven Facial Animation

TL;DR

This work tackles the coupling of near-homophonic syllables in self-supervised audio encoders used for 3D speech-driven facial animation. It introduces Wav2Sem, a plug-and-play semantic decorrelation module that extracts global audio semantics and aligns them to a sentence-level semantic space, mitigating phoneme-level entanglement. By fusing a semantic vector with phoneme features through a simple FC-based fusion and optimizing with an $L_1$ loss on LibriSpeech, Wav2Sem improves lip-sync accuracy and facial dynamics across multiple baselines and datasets, and even enhances phoneme recognition. The results demonstrate substantial gains in lip vertex accuracy, upper-face dynamics, and perceptual quality, supporting the practical value of semantic decorrelation for robust, expressive speech-driven animation.

Abstract

In 3D speech-driven facial animation generation, existing methods commonly employ pre-trained self-supervised audio models as encoders. However, due to the prevalence of phonetically similar syllables with distinct lip shapes in language, these near-homophone syllables tend to exhibit significant coupling in self-supervised audio feature spaces, leading to the averaging effect in subsequent lip motion generation. To address this issue, this paper proposes a plug-and-play semantic decorrelation module-Wav2Sem. This module extracts semantic features corresponding to the entire audio sequence, leveraging the added semantic information to decorrelate audio encodings within the feature space, thereby achieving more expressive audio features. Extensive experiments across multiple Speech-driven models indicate that the Wav2Sem module effectively decouples audio features, significantly alleviating the averaging effect of phonetically similar syllables in lip shape generation, thereby enhancing the precision and naturalness of facial animations. Our source code is available at https://github.com/wslh852/Wav2Sem.git.

Figures (5)

• Figure 1: Our plug-and-play Wav2Sem module effectively alleviates the coupling of near-homophonic syllables in pre-trained self-supervised models by extracting semantic features from audio.
• Figure 2: Overview of Wav2Sem, which is trained on a large dataset of text-audio pairs. Given input audio signals, Wav2Sem extracts semantic information using a Temporal Convolutional Network (TCN) and a Transformer-based encoder. The learned semantic information is aligned with the text semantics in the BERT space. Integrating Wav2Sem into the audio encoder of an arbitrary speech-driven facial animation framework enables the decoupling of the near-homophonic syllable feature space, resulting in varying performance enhancement.
• Figure 3: T-SNE comparison for near-homophonic syllables.
• Figure 4: Evaluations of facial motions with and without Wav2Sem for different methods on VOCASET (left) and BIWI (right).
• Figure 5: Visualizations of different audio encoding structures