This repository includes two major projects focused on natural language processing and speech processing.
Built a Transformer model from scratch and trained it on the Canadian Hansards dataset for French-to-English translation. Key components include:
- Implementation of LayerNorm, Multi-Head Attention, and FeedForward layers
- Full encoder-decoder architecture with both pre- and post-layer normalization
- Greedy and Beam Search decoding algorithms
- BLEU score calculation for evaluation
- Model training, testing, and comparison with pre-trained and commercial translation systems
Developed and analyzed several machine learning and dynamic programming techniques for speech data. Includes:
- Speaker identification using Gaussian Mixture Models
- Deception detection with GRU-based models
- Word Error Rate evaluation using Levenshtein distance
- Speaker verification using Dynamic Time Warping (DTW)
- Empirical analysis and performance discussions
All code is written in Python 3 and tested on the teach.cs environment.