HuBERT + ViViT dual encoders with bidirectional cross-modal attention on RAVDESS corpus. Walrus Optimizer for post-training feature selection on 1,536-dimensional joint embedding. Evaluated across 4-SNR robustness levels (15, 10, 5, 0 dB) with MUSAN noise corpus. Led architectural decisions and authored the complete paper.
Dual-Transformer Cross-Attention Multimodal Emotion Recognition
2026 · IIT Roorkee · Electrical Engineering Department · Co-author