SinglishVoice: Code-mixed Romanized Sinhala Text-to-Speech System

## Overview

SinglishVoice is a research project and implementation of a Text-to-Speech (TTS) system designed to handle code-mixed Romanized Sinhala (Singlish).
Unlike traditional Sinhala TTS systems that require native Sinhala script, SinglishVoice processes Romanized input text — including informal spelling, shorthand, and code-mixed English — and produces natural Sinhala speech.

This project was developed as part of the BEng in Software Engineering dissertation at the University of Westminster in collaboration with the Informatics Institute of Technology.

---

Sinhala speakers frequently use Romanized Sinhala on social media and digital platforms.
However, existing TTS systems do not support Romanized Sinhala input, creating accessibility challenges for users.

SinglishVoice bridges this gap by introducing an end-to-end TTS pipeline:

1. Back-Transliteration: Romanized Sinhala → Native Sinhala (using a fine-tuned NLLB model).
2. Speech Synthesis: Native Sinhala → Natural Sinhala speech (using a VITS model).

---

Features

• Support for code-mixed Romanized Sinhala (Singlish).
• Back-transliteration using fine-tuned NLLB.
• High-quality speech synthesis with VITS (Vocoder-free TTS).
• User-friendly GUI for text input and speech generation.
• Modular and scalable architecture.

---

Evaluation Results

• NLLB Model (Transliteration)

  • Word Error Rate (WER): 21%
  • Character Error Rate (CER): 6%
  • BLEU Score: 58%

• VITS Model (Text-to-Speech)

  • Male Speaker MOS: 3.8
  • Female Speaker MOS: 2.6

Note: Code-mixed sentences and female voice synthesis remain challenging.

---

System Architecture

1. Input: Romanized Sinhala text (Singlish).
2. Back-Transliteration Module (NLLB).
3. Speech Synthesis Module (VITS).
4. Output: Natural Sinhala speech.

---

Tech Stack

• Languages: Python
• Frameworks & Libraries:
  • PyTorch
  • Hugging Face Transformers (NLLB)
  • Coqui TTS (VITS)
• Datasets:
  • Swa Bhasha Dataset
  • Dakshina Dataset
  • PathNirvana (7.5 hrs audio)

---

Installation & Usage

# Clone repository
git clone https://github.com/your-username/singlishvoice.git
cd singlishvoice

# Run Backend
cd Backend && pip install -r requirements.txt && uvicorn main:app --host 0.0.0.0 --port 8080

# Run Frontend
cd Frontend && pip install -r requirements.txt && streamlit run app.py

---

Use Cases

• Accessibility support for visually impaired users.
• Social media content consumption in Romanized Sinhala.
• Language learning & research in Sinhala NLP.
• Enabling voice-enabled Sinhala applications.

---

Future Work

• Improve female voice naturalness.
• Enhance handling of complex code-mixed sentences.
• Optimize system for real-time mobile applications.

---

Author

Chamika Uluwatta
BEng in Software Engineering, University of Westminster

Supervised by Dr. Ruvan Weerasinghe

---