Comparing Model-Centric and Data-Centric Strategies for Italian Visual Speech Recognition in Low-Resource Scenarios

Visual Speech Recognition (VSR) enables speech understanding without relying on acoustic signals, making it particularly valuable for assistive applications. However, its development in Italian is limited by the lack of large-scale annotated datasets. In this work, we present a comparative study of two strategies to improve a LipNet-based baseline trained on VSR-IT19, a novel task-oriented dataset composed of custom sentences recorded by Italian speakers for assistive communication. First, we investigate a model-centric approach based on Online Knowledge Distillation (OKD) with a dynamic attention mechanism, reducing the Word Error Rate (WER) from 45.13% to 42.00%. Second, we explore a data-centric strategy based on synthetic data augmentation using Wav2Lip, further reducing the WER to 38.00%. Results show that both approaches are effective, with synthetic augmentation providing the largest improvement, highlighting the importance of data diversity in low-resource VSR settings.