RAVURI VANUSSHKA

ManoDhwani: Reliability-Gated Multimodal Fusion for Depression Detection Using RoBERTa and WavLM

Developed a multimodal deep learning system for automatic depression detection using textual and audio features from the DAIC-WOZ dataset. Integrated RoBERTa for text representation and WavLM for speech representation, combining both modalities through cross-modal attention to improve depression classification. Explored techniques such as focal loss and label smoothing to support model training and classification performance.

MT2KD – General-Purpose Audio Encoder Using Knowledge Distillation

Implementing MT2KD, a research-based audio representation learning framework that uses knowledge distillation to transfer knowledge from specialized teacher models into a unified student encoder. The project explores audio tagging, automatic speech recognition, and speaker verification using teacher models such as BEATs, Whisper, and ECAPA-TDNN. Currently working on the audio-tagging pipeline using the FSD50K dataset, including data preprocessing, teacher–student learning, and model evaluation.

UrbanEye AI – AI-Powered Urban Traffic Intelligence System

Developed UrbanEye AI, an intelligent traffic monitoring system that transforms existing CCTV and ANPR cameras into urban traffic intelligence infrastructure. The system integrates vehicle and license plate detection, automatic number plate recognition using OCR, and multi-camera trajectory tracking to analyze vehicle movement across camera locations. The proposed platform aims to support traffic-flow analysis, congestion monitoring, vehicle tracking, and GIS-based traffic visualization.