PsycheTTS: Emotional Text-to-Speech Platform

University of Sri Jayewardenepura Research Contribution

An advanced research initiative focused on developing high-fidelity, emotionally expressive Text-to-Speech models for low-resource languages (Sinhala).

View Research Portal • Data Collector

---

🚀 Project Overview

PsycheTTS combines the power of Large Language Models (Gemini 2.5 Pro) with efficient neural architectures (Bi-Directional Mamba) to solve the data scarcity problem in Sinhala TTS.

Key Components

Module	Description	Status
Collector	A Flask-based tool to generate & curate emotional audio samples using Gemini API. Features a Manual Helper for precise control.	✅ Verified Working
Research	The core training pipeline implementing the PsycheTTS Architecture (Style Encoder + Mamba + Flow Matching).	🚧 In Development
Portal	Interactive HTML dashboard for visualizing architecture, pipelines, and results.	✨ Live

---

📂 Directory Structure

PsycheTTS/
├── collector/          # Data Collection Suite
│   ├── dataset/        # Validated Audio Samples (.wav) & Metadata (.csv)
│   ├── src/web/        # Web Interface (Flask)
│   └── src/cli/        # Command Line Tools
│
├── research/           # Model Training & Experiments
│   ├── psyche_tts_portal.html  # Visualization Dashboard
│   ├── train.py        # Main Training Loop
│   ├── model.py        # Architecture Definitions
│   └── modules.py      # Core Building Blocks (Mamba, etc.)

---

Why use Gemini 2.5 Pro preview-tts?

Due to the scarcity of Sinhala voice samples with expected quality, reliability, and consistency, we had to find another solution for the voice dataset. After our research and comparison, we figured out that Gemini has very good emotional and natural-sounding vocal ability in Sinhala TTS generation.

⚡ Quick Start

1. Data Collector (Start Here)

Generate your own emotional dataset using our specialized tool.

# 1. Navigate to the collector directory
cd collector

# 2. Install dependencies (First time only)
pip install -r requirements.txt

# 3. Launch the Server
python src/web/server.py

Note: The server will start at http://127.0.0.1:8000/manual. Ensure you have your GEMINI_API_KEY ready if using automated features.

2. Research & Training

Train the model on the collected data.

# 1. Navigate to research directory
cd research

# 2. Install dependencies
pip install -r requirements.txt

# 3. Start Training
python train.py

---

🤝 Collaboration

This project is open for collaboration, specifically targeting researchers from the University of Sri Jayewardenepura.

• Dataset Contributions: Please ensure all new samples are logged in collector/dataset/metadata.csv.
• Model Improvements: Submit PRs for new modules in research/modules.py.

---

_{© 2025 University of Sri Jayewardenepura & Cortana R&D Contribution. All Rights Reserved.}