M.Tech Computer Science Engineering - Artificial Intelligence at NSUT (Class of 2024)

Shivangi Singh  • Peerlist

Implemented a real-time gaze detector using OpenCV and Dlib, enabling accurate eye tracking through a webcam. Utilized facial landmark detection and eye aspect ratio calculations to monitor and analyze gaze direction in real-time.

REAL-TIME GAZE DETECTOR

Developed a text-to-SQL application leveraging Gemini AI Pro, enabling prompt-based query generation executed against SQL databases for data retrieval.

LLM BASED TEXT TO SQL APPLICATION

This project aims at understanding and mitigating customer attrition from credit card services. Utilizing SQL queries and Python analysis to explore data variables. Set up database on AWS RDS for real-world data scenarios. Visualized the results on Tableau dashboard highlighting key factors in predicting and preventing churn, incorporating insights from the analysis.

CUSTOMER CHURN ANALYSIS

Developed a personalized content - based movie recommendation system which recommends movies similar to user’s choices and analyzes reviews scraped from IMDB website for enhanced recommendations.

NETFLIX RECOMMEDATION SYSTEM

Developed a question answering system using a fine-tuned BERT model from Hugging Face Transformers, focused on answering queries based on the CoQA dataset. Responsibilities included handling data preprocessing, model integration, and performance optimization.

QUESTION ANSWERING WITH FINE TUNED BERT MODEL

Automatic speech recognition (ASR) stands out as a pivotal technique, steadily gaining prominence by facilitating seamless communication between humans and technology. Due to a shortage of substantial speech datasets for model training, ASR systems for most languages lack acceptable performance, despite achieving human-like performance for a limited number of languages. In this research project, Hugging Face ecosystem is leveraged to fine-tune state-of-the-art Whisper Speech architecture to showcase the efficacy of transfer learning technique in ASR tasks, yielding crucial enhancements in accuracy and performance compared to conventional training methods. The findings underscore the potential of transfer learning to advance ASR technologies and provide insights for future research endeavors in this dynamic domain.&nbsp;

Deep Transfer Learning based Hindi Speech Recognition

A phoneme represents the smallest distinct sound unit within a language, unique to each linguistic system. Phoneme identification finds applications in various speech-related technologies, including automatic speech recognition, language learning, assistive technology for speech impaired individuals. The realm of Automatic Speech Recognition (ASR) stands as a formidable challenge within the field, with a predominant emphasis on extensively spoken languages. However, the intricacies posed by Hindi speech recognition have received limited attention, owing to the language's complex structure and the scarcity of pertinent data. Existing models designed for languages like English cannot be seamlessly adapted to the nuances of Hindi. This research project presents an efficient approach to phoneme-based speech recognition in the Hindi language through the implementation of a deep learning model. The proposed model leverages techniques in ASR to transform raw audio data into phoneme sequences, assisting in recognizing digits from a dataset collected in limited capacity consisting of counting in Hindi language.