license: mit
datasets:
- sameerbanchhor/CHATTISGARHI-TTS-F
language:
- hi
pipeline_tag: text-to-speech
tags:
- chattisgarhi
- chhattigarh
Chhattisgarhi Text-to-Speech (TTS) Model — VITS Based
A deep learning-based text-to-speech (TTS) model for the Chhattisgarhi language, trained using the VITS architecture. This project is designed to make technology more accessible for the people of Chhattisgarh by converting Chhattisgarhi text into natural, regional speech.
Author: Sameer Banchhor
MSc Student at Hemchand Yadav University
📧 Email: [email protected]
🐙 GitHub: @sameer-banchhor-git
🔗 LinkedIn: Sameer Banchhor
🌟 Project Highlights
- 🔊 Language Support: Chhattisgarhi (regional dialects and tones)
- 🧠 Model Architecture: VITS (Variational Inference Text-to-Speech)
- 🎯 Goal: Enable high-quality speech generation for educational, informational, and accessibility applications in Chhattisgarh
- 🛠️ Current Status: Model trained and functional; app development planned
📚 Dataset
Primary Dataset:
CHATTISGARHI-TTS-F on Hugging Face — curated and released by the author
Data Composition:
- Regional sentences in Chhattisgarhi, including various tones and expressions
- Transcriptions aligned with high-quality audio
- Data sourced from:
- YouTube spoken content (sentence-transcription pairs)
- IISE dataset for phonetic richness and clarity
Sample Sentences with Regional Tone:
- "राजस्थान के नामी ब्यंजन चूरमालाड़ू गुड़ के पाग म गहूँ के दरदरहा पिसान के लाड़ू म तिली अउ नरियल के सुवाद म सजथे"
- "दुग्ध क्रान्ति भारत के योजना हे जेखर ले भारत म दूध के कमी ला दुरिहा करे जा सकथे एला श्वेत क्रांति घलोक कहिथे"
- "जम्मू कश्मीर म पर्यटन उद्योग ला बढ़ावा देना उहाँ के अर्थबेवस्था ला सुचारू रूप ले चलाय बर जरुरी हे"
⚙️ Model Training
| Component | Details |
|---|---|
| Architecture | VITS (with adversarial training) |
| Framework | PyTorch |
| Audio Sampling | 22050 Hz (standard TTS rate) |
| Dataset Size | 27 GB |
| Optimizer | Adam |
| GPU | NVIDIA RTX 3090 |
| Vocoder Used | Integrated with VITS |
| Text Normalization | Custom normalization for Devanagari |
🧪 Inference Example
from models import Synthesizer
tts = Synthesizer(
checkpoint_path="models/chhattisgarhi_vits.pth",
config_path="configs/config.json"
)
tts.speak("जम्मू कश्मीर म पर्यटन उद्योग ला बढ़ावा देना उहाँ के अर्थबेवस्था ला सुचारू रूप ले चलाय बर जरुरी हे")
🛣️ Roadmap
- ✅ Train and test baseline VITS model on Chhattisgarhi dataset
- 🔄 Add support for tonal variation and emphasis
- 🧹 Improve preprocessing for low-resource sentence cleaning
- 📱 Develop full-featured TTS application (mobile/desktop/web)
- 🌐 Incorporate transliteration (for users who can’t type in Devanagari)
📄 License
MIT License (recommended for open-source contributions)
🤝 Contributing
If you're a linguist, developer, or enthusiast in low-resource languages and speech tech, contributions are welcome! Feel free to fork the repo or reach out directly.
🙋 Support & Contact
For any issues, suggestions, or collaborations:
📧 Email: [email protected] 📍 GitHub: github.com/sameer-banchhor-git