Show HN: Open dataset for distilling GPU audio2face models to CPU students (huggingface.co)

A new open dataset has been released aimed at distilling GPU-bound Audio2Face models into CPU-friendly versions, enabling real-time performance on less powerful hardware. Created by myned-ai, this dataset comprises 14,082 annotated emotional-speech clips, specifically designed for research on teacher-student model distillation. It features intricate details such as ARKit blendshape sequences derived from NVIDIA’s Audio2Face and Audio2Emotion models, along with emotional conditioning vectors. Researchers can access the dataset to explore the behavior of these models and improve efficiency in rendering 3D facial animations driven by audio input. This release holds significant implications for the AI/ML community, particularly for those working in the fields of emotion recognition and synthetic animation. By focusing on specific actor styles rather than generic representations, the dataset allows for targeted research into optimizing model architectures while still leveraging complex emotional cues. Moreover, the dataset's structure facilitates integration with existing audio sources from several public corpora, providing a flexible base for future studies and applications in real-time emotional expression synthesis in virtual environments.