Grapheme-to-Phoneme Conversion using Encoder-Decoder LSTMs

This project implements sequence-to-sequence encoder-decoder architectures for Grapheme-to-Phoneme (G2P) conversion using PyTorch.

The project compares three sequence modelling approaches:

1. Bottleneck Encoder-Decoder
2. Fixed Context Vector Decoder
3. Attention-Based Decoder

The models are trained on the CMU Pronouncing Dictionary dataset (CMUdict).

---

Key Concepts

• Sequence-to-Sequence Learning
• LSTMs from Scratch
• Encoder-Decoder Architectures
• Attention Mechanisms
• Cross-Attention
• Speech and Language Processing
• Grapheme-to-Phoneme Conversion

---

Features

• Custom LSTM implementation (without nn.LSTM)
• Dot-product attention implementation
• Greedy decoding
• Hyperparameter tuning
• Phoneme Error Rate (PER) evaluation
• Attention heatmap visualisation

---

Technologies Used

• Python
• PyTorch
• NumPy
• Matplotlib
• Pandas

---

Results

The attention-based encoder-decoder achieved the best performance by overcoming the hidden-state bottleneck problem.

Key findings:

• Attention improved convergence speed
• Lower Phoneme Error Rate (PER)
• Better handling of longer words
• Improved sequence alignment

---

Repository Structure

• notebooks/ → Main notebook
• report/ → final report

---

Author

Kamva