CoolFace
Modelpublic

Satyam-Srivastava/TTS_haryanvi_bangru_vits_based_model

sourceHugging Faceapache-2.0updated 5mo agoView on Hugging Face
0likes9downloads
Model Card

language:

  • —haryanvi(bangru) tags:
  • —text-to-speech
  • —TTS
  • —generative-ai
  • —speech
  • —low-resource-language
  • —haryanvi license: apache-2.0 datasets:
  • —Satyam-Srivastava/RDS metrics:
  • —mos
  • —wer pipeline_tag: text-to-speech ---

🎤 Haryanvi Text-to-Speech (TTS) Model

This model converts Haryanvi text into natural-sounding speech.

It is designed to support low-resource Indian dialects, focusing on:

  • —Native pronunciation
  • —Regional tone
  • —Real-world usability (education, voice assistants)

⚡ Built as part of Data Science & AI Lab Project.

🚀 Motivation

Most TTS systems do not support regional dialects like Haryanvi.

This project aims to:

  • —Preserve regional language identity
  • —Enable voice-based learning tools
  • —Build inclusive AI for Indian users

💡 Unlike generic Hindi TTS, this model captures Haryanvi phonetics and accent patterns.

🧠 Model Details

  • —Architecture: VITS
  • —Vocoder: HiFi-GAN
  • —Framework: PyTorch
  • —Task: Text-to-Speech (TTS)

Input

  • —Haryanvi text

Output

  • —Speech waveform (.wav)

📊 Dataset

  • —Satyam-Srivastava/RDS(audiodata is copied from ankitdhiman_you can see in card as well. For audio data credit goes to Ankitdhiman.)
  • —Total Samples: 5515
  • —Duration: approx 1 hours
  • —Format: WAV files

Preprocessing

  • —Noise removal
  • —Normalization
  • —Text cleaning (Hindi → Haryanvi phonetics)

⚙️ Training

  • —Epochs: 100
  • —Batch Size: 32
  • —Optimizer: ---
  • —Learning Rate: 1e-5

Hardware

  • —GPU: (T4 / A100 )

Training Strategy

  • —Fine-tuning on low-resource data
  • —Augmentation for robustness

📈 Evaluation

Metrics

  • —MOS (Mean Opinion Score): X.X
  • —WER (Word Error Rate): X%

Observations

  • —Good pronunciation for common phrases
  • —Slight errors in rare words

🎧 Sample Outputs

TextAudio
"रै, काल रात नै इतणी तेज आँधी आई कि सारा गाँव धूल तै भर ग्या अर किम्मे भी साफ नी दिख रया था"Play Audio
"अर सुन, जिब खेतां मं पानी भर जावै सै ना, तै बैल भी चालण तै कतरावै सै अर किसान परेशान हो जावै सै"Play Audio
"यो छोरा सारा दिन मोबाइल मं घुस्या रहवै सै, ना घर का काम देखै सै अर ना पढ़ाई की कोई फिकर करै सै"Play Audio
"जिब तै बरसात शुरू होई सै, तब तै गलियां मं कीचड़ हो रया सै अर आवण-जावण मं भी बड़ी दिक्कत आवै सै"Play Audio
"रै, या दुनिया भी बड़ी अजीब सै, जिब जरूरत हो सै तब कोई साथ नी देवे, अर जिब सब ठीक हो जावै सै तै सब नेड़े आ जावै सै"Play Audio
"घर मं जिब बिजली चली जावै सै ना, तै सबने पंखा याद आवै सै अर कोई भी चैन तै बैठ नी सकै"Play Audio
"यो कुत्ता भी बड़ा चालाक सै, जिब देखै सै कोई ना सै तै चुपके तै रोटी उठा के भाग जावै सै"Play Audio
"जिब बूढ़े लोग आपणी पुराणी बातें सुनावै सै ना, तै लागै सै जैसे टाइम पीछे चल ग्या हो अर सब कुछ आंखां के आगे हो रया हो"Play Audio

⚠️ Limitations

  • —Limited dataset size

🔮 Future Improvements

  • —Multi-speaker Haryanvi TTS
  • —Emotion-aware speech generation
  • —Integration with RAG-based learning platform

🧪 Usage

python
from transformers import pipeline

tts = pipeline("text-to-speech", model="your-username/model-name")

output = tts("राम राम भाई")

👨‍💻 12. Contributors

markdown
## 👨‍💻 Contributors

- Satyam – Documentation & Model run 
- Abhishek – Text Processing 
- Fazlur – Audio Pipeline
- Sanket – Deployment & HuggingFace


## ❤️ Acknowledgements

Thanks to:
- Hugging Face
- Open-source TTS community
- Course instructors & TAs

## 🌍 Impact

This project contributes to:
- Digital inclusion of rural dialects
- Voice-based education systems
- Preservation of regional languages