TTS_Code/final_model/README.md

# 🗣️ Menstrual Bangla TTS

## 📌 Model Overview
This model is a LoRA-adapted version of **VoxCPM** designed for:

- Bangla speech synthesis  
- Domain-specific menstrual health narration  
- Clean and natural TTS output  
- Lightweight inference  

**Base Model:** VoxCPM-0.5B  
**Fine-tuning:** LoRA  
**Language:** Bangla  

---

## 🚀 Installation

Install required dependencies:

pip install torch soundfile transformers huggingface_hub

---

## 🔑 Authentication (Required for Private Models)

If the repository is private, authenticate using:

huggingface-cli login  

OR in Python:

from huggingface_hub import login  
login()

Use your own Hugging Face access token.

---

## 💻 Usage Example

import sys  
from huggingface_hub import snapshot_download  
from pathlib import Path  
import soundfile as sf  

# Replace with your repo path if under organization  
repo_id = "SibgatUl/menstrual_bangla_tts"  

# Download model files  
local_dir = snapshot_download(repo_id=repo_id)  
print("Downloaded to:", local_dir)  

ROOT = Path(local_dir)  

# Add local model paths  
sys.path.insert(0, str(ROOT))  
sys.path.insert(0, str(ROOT / "VoxCPM"))  

# Import loader  
from load_model import load_menstrual_bangla_tts  

# Load model  
model = load_menstrual_bangla_tts(str(ROOT))  

# Generate audio  
audio = model.generate(  
    text="এটি একটি পরীক্ষা",  
    cfg_value=2.0,  
    inference_timesteps=10,  
)  

# Save output  
sf.write("output.wav", audio, model.tts_model.sample_rate)  

print("Generation successful.")

---

## ⚙️ Generation Parameters

text — Bangla text input  
cfg_value — Guidance strength (default ~2.0)  
inference_timesteps — Number of diffusion steps (higher = better quality but slower)  

Recommended configuration:

cfg_value = 2.0  
inference_timesteps = 10–20  

---

## 👤 Author

Developed by **Maisha Rahman**  
Affiliation: Machine Intelligence Lab