Grapheme-to-Phoneme (G2P) Conversion

Implementation of a sequence-to-sequence model with attention mechanism for Grapheme-to-Phoneme (G2P) conversion — the task of converting written text (graphemes) to their phonetic representations (phonemes).

---

Introduction

Grapheme-to-Phoneme conversion is a fundamental task in speech processing, essential for:

• Text-to-Speech (TTS) Systems: Converting text input to pronunciation for speech synthesis
• Automatic Speech Recognition: Mapping recognized phonemes back to text
• Pronunciation Dictionaries: Generating phonetic transcriptions for new words

---

Architecture

Encoder-Decoder with Attention

The model uses a classic sequence-to-sequence architecture:

Input: "hello" (graphemes: h-e-l-l-o)
  ↓
Encoder LSTM → Hidden States
  ↓
Bahdanau Attention
  ↓
Decoder LSTM → Output Sequence
  ↓
Output: "HH AH0 L OW1" (phonemes in ARPAbet)

Key Components

1. Bidirectional Encoder: Captures context from both directions of the input sequence
2. Bahdanau Attention: Learns to focus on relevant input characters for each output phoneme
3. Decoder with Teacher Forcing: Generates phoneme sequence with training stabilization

---

Features

• Clean, modular PyTorch implementation
• Training scripts with configurable hyperparameters

---

Results

None	Accuracy (Token)	Accuracy (Word)	Word Error Rate
Training Dataset	99.74%	94.09%	0.82%
Evaluate (Dev) Dataset	99.52%	93.77%	0.93%

---

Applications

This G2P model can be integrated into:

• Custom TTS pipelines
• Pronunciation learning apps
• Linguistic research tools

---

Repository

Code and pretrained models: https://github.com/QuyAnh2005/Grapheme-To-Phoneme