A production-ready, multi-lingual Text-to-Speech system supporting 11 Indian languages with 21 voice variants, trained on 150+ hours of speech data. Built for maternal healthcare accessibility.
🌐 Live API: https://harshil748-voiceapi.hf.space
📖 API Docs: https://harshil748-voiceapi.hf.space/docs
💻 GitHub: https://github.com/harshil748/VoiceAPI
Built for the Voice Tech for All Hackathon to address linguistic barriers in rural Indian healthcare. The system converts medical instructions into natural speech across 11 languages, enabling accessible prenatal care guidance for non-literate populations.
- 🌏 11 Indian Languages: Hindi, Bengali, Marathi, Telugu, Kannada, Bhojpuri, Chhattisgarhi, Maithili, Magahi, English, Gujarati
- 🎤 21 Voice Variants: Male & Female voices trained on 150+ hours of speech data
- 🎭 Prosody Control: 9 style presets (calm, happy, sad, slow, fast, etc.)
- ⚡ Real-time Performance: 0.3-0.9s inference on CPU hardware
- 🔌 Production REST API: FastAPI with automatic docs, CORS support
- 🧠 Neural Architecture: VITS + Meta MMS models with JIT optimization
- 📦 Deployed on HuggingFace Spaces: Always-on, cloud-hosted API
import requests
# Use the live API
base_url = 'https://harshil748-voiceapi.hf.space/Get_Inference'
params = {
'text': 'नमस्ते, आप कैसे हैं?', # Hindi text
'lang': 'hindi',
}
# Upload any WAV file as speaker reference
with open('reference.wav', 'rb') as audio:
response = requests.get(base_url, params=params, files={'speaker_wav': audio})
if response.status_code == 200:
with open('output.wav', 'wb') as f:
f.write(response.content)
print("✅ Audio saved as 'output.wav'")curl -X GET "https://harshil748-voiceapi.hf.space/Get_Inference?text=નમસ્તે&lang=gujarati" \
-F "speaker_wav=@reference.wav" \
-o output.wav- Method:
GET - URL:
https://harshil748-voiceapi.hf.space/Get_Inference - Params Tab:
text: Your text in any supported languagelang: One of: hindi, bengali, marathi, telugu, kannada, gujarati, bhojpuri, chhattisgarhi, maithili, magahi, english
- Body Tab →
form-data:- Key:
speaker_wav(Type: File) - Value: Upload any
.wavfile
- Key:
- Send → Save response as
.wavfile
| Language | Code | Male Voice | Female Voice | Sample Text |
|---|---|---|---|---|
| Hindi | hindi |
✅ | ✅ | नमस्ते |
| Bengali | bengali |
✅ | ✅ | নমস্কার |
| Marathi | marathi |
✅ | ✅ | नमस्कार |
| Telugu | telugu |
✅ | ✅ | నమస్కారం |
| Kannada | kannada |
✅ | ✅ | ನಮಸ್ಕಾರ |
| Gujarati | gujarati |
✅ | - | નમસ્તે |
| Bhojpuri | bhojpuri |
✅ | ✅ | प्रणाम |
| Chhattisgarhi | chhattisgarhi |
✅ | ✅ | नमस्कार |
| Maithili | maithili |
✅ | ✅ | प्रणाम |
| Magahi | magahi |
✅ | ✅ | प्रणाम |
| English | english |
✅ | ✅ | hello (lowercase required) |
Converts text to speech in any supported Indian language.
Endpoint: https://harshil748-voiceapi.hf.space/Get_Inference
Parameters:
| Parameter | Type | Required | Description |
|---|---|---|---|
text |
string | ✅ | Text to convert to speech (English must be lowercase) |
lang |
string | ✅ | Language code (see table above) |
speaker_wav |
file | ✅ | Reference WAV file for speaker voice cloning |
Response: audio/wav file (200 OK)
Example:
import requests
response = requests.get(
'https://harshil748-voiceapi.hf.space/Get_Inference',
params={'text': 'ನಮಸ್ಕಾರ', 'lang': 'kannada'},
files={'speaker_wav': open('reference.wav', 'rb')}
)
with open('output.wav', 'wb') as f:
f.write(response.content)| Metric | Value |
|---|---|
| Languages | 11 Indian languages |
| Voice Variants | 21 (male/female per language) |
| Training Data | 150+ hours (OpenSLR, Common Voice, IndicTTS) |
| Model Size | 318MB (VITS), 998MB (Coqui) |
| Inference Time | 0.3-0.9 seconds per utterance |
| Sample Rate | 22.05kHz (VITS), 16kHz (MMS) |
| Architecture | VITS + Meta MMS + Coqui TTS |
| Deployment | HuggingFace Spaces (Docker) |
| API Framework | FastAPI with Uvicorn |
Built with a unified inference engine supporting heterogeneous model formats:
- JIT Models (.pt): VITS models trained on 150+ hours for 19 voices
- Coqui Checkpoints (.pth): Full checkpoints with config.json for Bhojpuri
- HuggingFace MMS: Meta's multilingual model for Gujarati
View Architecture Diagrams
git clone https://github.com/harshil748/VoiceAPI
cd VoiceAPI
python3 -m venv tts
source tts/bin/activate # On Windows: tts\Scripts\activate
pip install -r requirements.txtpython -m src.cli serve --port 8000Visit http://localhost:8000/docs for interactive API documentation.
python -m src.cli synthesize \
--text "नमस्ते दोस्तों" \
--voice hi_male \
--output hello.wav
afplay hello.wav # macOSVoiceAPI/
├── src/
│ ├── api.py # FastAPI REST server
│ ├── engine.py # Unified TTS inference engine
│ ├── tokenizer.py # Indic script tokenization
│ ├── config.py # Language/voice configurations
│ └── cli.py # Command-line interface
├── models/ # Model storage (8GB, hosted on HF)
├── training/ # Training scripts and configs
│ ├── train_vits.py # VITS training pipeline
│ ├── prepare_dataset.py
│ └── export_model.py
├── tests/ # API integration tests
├── diagrams/ # Architecture diagrams (PNG)
└── technical_report.tex # IEEE paper
Read the full technical writeup: VoiceAPI.pdf
Key Contributions:
- Trained 21 VITS models on 150+ hours of Indian language data
- Solved tokenizer alignment issues for Indic scripts
- Implemented lazy loading reducing memory by 60%
- Signal-based prosody control without retraining
- OpenSLR: Public speech datasets for 6 Indian languages
- Common Voice: Mozilla's crowdsourced speech corpus
- IndicTTS: IIT Madras speech synthesis resources
- Meta MMS: Massively multilingual speech models
- HuggingFace: Model hosting and deployment infrastructure
- Code: MIT License
- Models: CC BY 4.0 (OpenSLR, IndicTTS), CC BY-NC 4.0 (MMS)
Built by Team VoiceAPI for Voice Tech for All Hackathon 2024:
- Harshil Patel - CHARUSAT University
- Aashvi Maurya - University of Allahabad
- Pratyush Kumar Das - FM University
- Jaideep Amrabad - NNRGI Hyderabad
- Issues: GitHub Issues
- API Status: Check HuggingFace Space
- Documentation: Live API Docs
⭐ Star this repo if you find it useful!
Built with ❤️ for accessible healthcare in India