Attentional Multimodal Speech Enhancement for Voice User Interface in Consumer Electronics

Voice User Interfaces (VUI) in consumer electronics often operate in noisy environments where speech quality degrades significantly. We propose AV-Net, a lightweight audiovisual speech enhancement model that addresses this challenge through novel cross-attentional feature fusion. Our approach dynamically integrates audio and visual modalities using a computationally efficient architecture combining a convolutional encoder-decoder (5.2M parameters) with a P3D-ResNet18 video encoder. The key innovation is a cross-attention mechanism that learns inter-modal correlations while preserving modality-specific features, outperforming conventional fusion methods. Evaluated on TCD-TIMIT and AVSE3 datasets under challenging conditions (SNR \leq -5dB), AV-Net achieves a PESQ of 2.56 (vs. 1.26 baseline) and STOI improvement of 22%, while maintaining real-time performance (RTF=0.11). The model demonstrates strong generalization to unseen speakers and diverse noise types, making it particularly suitable for resource-constrained edge devices in healthcare, automotive, and smart home applications where robust speech interaction is critical.