Pratham Gupta
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Pratham Gupta has a total of 4 co-authors. Here are the top co-authors by number of shared publications:
- Lakshya Gupta
Dept. of Mechanical Engineering IIT Kanpur
1 shared publication - Praveen Kumar
Department of Electrical Engineering, Perception and Intelligence Lab, Indian Institute of Technology Kanpur, Kanpur, India
1 shared publication Shubham Yadav
Dept. of Mechanical Engineering IIT Kanpur
1 shared publication- Tushar Sandhan
Department of Electrical Engineering, Perception and Intelligence Lab, Indian Institute of Technology Kanpur, Kanpur, India
1 shared publication
Latest Publications
- Open Access
- 29 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in TechRxiv
CAFNet: Cross Attentive Face-Voice Fusion Network
We present CAFNet, a cross-attentive fusion network that aligns and fuses speaker and face representations for robust face-voice verification in multilingual settings. Our pipeline combines audio preprocessing with strong modality, specific embeddings and a bidirectional cross-attention fusion head trained with cross-entropy. We report progress-phase results on the English-Urdu dataset of MAV-Celeb: an initial WavLMbased pipeline yielded an overall EER ≈ 0.2702, while the final ECAPA-TDNN + CAFNet pipeline improved this to 0.26316. The proposed architecture consistently improves cross-modal alignment and is compact enough to train on limited data.
- 29 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in TechRxiv
Open AccessCAFNet: Cross Attentive Face-Voice Fusion Network
We present CAFNet, a cross-attentive fusion network that aligns and fuses speaker and face representations for robust face-voice verification in multilingual settings. Our pipeline combines audio preprocessing with strong modality, specific embeddings and a bidirectional cross-attention fusion head trained with cross-entropy. We report progress-phase results on the English-Urdu dataset of MAV-Celeb: an initial WavLMbased pipeline yielded an overall EER ≈ 0.2702, while the final ECAPA-TDNN + CAFNet pipeline improved this to 0.26316. The proposed architecture consistently improves cross-modal alignment and is compact enough to train on limited data.