Work in computer vision, deep learning, and low-resource AI — centered on multilingual visual speech recognition.
Developing a language-agnostic system for real-time multilingual lip reading, addressing challenges in language dependency and adaptability in low-resource settings — including Bangla, which has almost no labeled VSR training data despite being the 7th most spoken language in the world. Applications include accessibility for deaf and hard-of-hearing communities, noisy-environment speech recognition, and multimodal communication systems.
Read the full write-upDeveloped and evaluated machine learning models for detecting Atrial Fibrillation (AF) from ECG signals using patient-aware data splitting. Implemented both a feature-based Random Forest classifier and a 1D Convolutional Neural Network directly on raw ECG segments, evaluated with accuracy and AUC. Built with Python (wfdb, scikit-learn, TensorFlow).
Utilized the LipNet architecture to train a lip reading model from video footage, curating the GRID Corpus Dataset for training and validation. Built a Tkinter-based interface for user video input and live prediction display — an early, single-language precursor to the current multilingual VSR thesis work.
Words by Animesh Bhattacharjee
animesh.tech