A collection of deep learning systems spanning speech, vision, and language — built from the ground up, including a from-scratch NumPy deep learning library used to train the final Transformer-based speech recognizer.
| Project | Description | Key result | Tech |
|---|---|---|---|
| Frame-Level Speech Phoneme Classification | Diamond-shaped MLP classifying speech frames into 42 phonemes from MFCC features | 88.44% validation accuracy | PyTorch, torchaudio |
| Face Classification & Verification | From-scratch ResNet-34 with ArcFace loss for closed-set classification and open-set verification | ~2.16% verification EER | PyTorch, ResNet, ArcFace |
| CTC Speech Recognition | CNN + pyramidal BiLSTM encoder mapping speech directly to phoneme sequences via CTC loss | 4.39 validation edit distance | PyTorch, CTC, BiLSTM |
| Deep Learning Library & Attention-Based ASR | NumPy deep learning library built from scratch (autograd through Transformers), used to train an attention-based speech recognizer | 8.78% character error rate | NumPy, PyTorch, Transformers |
Each project folder contains its own README with the problem statement, architecture, training setup, results, and key engineering decisions. The library project is the connective thread across all of them — the autograd, layer, and attention primitives implemented from scratch there are the same building blocks used, at the framework level, throughout the other projects.
Python · PyTorch · NumPy · torchaudio · Transformers / Attention · CTC · RNN / LSTM / GRU · CNN / ResNet · Mixed-precision training · Weights & Biases
Developed as part of a graduate deep learning curriculum covering neural network fundamentals through modern architectures, and revisited as a teaching assistant for the same material.