README.md

metadata

title: IndicConformer STT API
emoji: 🎙️
colorFrom: blue
colorTo: purple
sdk: docker
pinned: false
license: mit

IndicConformer Speech-to-Text API 🎙️

Fast and accurate Speech-to-Text API for 22 Indian languages powered by AI4Bharat's IndicConformer model.

🌟 Features

• 22 Indian Languages Supported: Hindi, Telugu, Bengali, Tamil, and 18 more
• Long Audio Support: Process up to 30 minutes of audio
• Parallel Processing: Fast transcription with chunked inference
• Multiple Formats: Supports WAV, MP3, FLAC, M4A

🚀 Quick Start

API Endpoints

• Base URL: Your Space URL
• Documentation: /docs (Interactive Swagger UI)
• Transcribe: POST /transcribe
• Health Check: GET /health

Example Usage

Using cURL

curl -X POST "https://your-space-url.hf.space/transcribe" \
  -F "file=@audio.wav" \
  -F "language=hi"

Using Python

import requests

url = "https://your-space-url.hf.space/transcribe"

files = {"file": open("audio.wav", "rb")}
data = {"language": "hi"}

response = requests.post(url, files=files, data=data)
print(response.json())

Using JavaScript

const formData = new FormData();
formData.append('file', audioFile);
formData.append('language', 'hi');

const response = await fetch('https://your-space-url.hf.space/transcribe', {
  method: 'POST',
  body: formData
});

const result = await response.json();
console.log(result.transcription);

🗣️ Supported Languages

Code	Language	Code	Language
hi	Hindi	te	Telugu
bn	Bengali	ta	Tamil
mr	Marathi	gu	Gujarati
kn	Kannada	ml	Malayalam
pa	Punjabi	or	Odia
as	Assamese	ur	Urdu
ne	Nepali	kok	Konkani
sd	Sindhi	doi	Dogri
brx	Bodo	mai	Maithili
mni	Manipuri	ks	Kashmiri
sa	Sanskrit	sat	Santali

📊 Response Format

{
  "success": true,
  "transcription": "आपका टेक्स्ट यहां",
  "metadata": {
    "audio_duration": 45.2,
    "audio_duration_minutes": 0.75,
    "inference_time": 2.1543,
    "rtf": 0.0476,
    "language": "hi",
    "decoder": "rnnt",
    "num_chunks": 2
  }
}

⚡ Performance

• Real-Time Factor (RTF): ~0.05 (20x faster than real-time on GPU)
• Max Audio Length: 30 minutes
• Chunk Processing: 30s chunks with 2s overlap for optimal accuracy

🛠️ Model Information

• Model: ai4bharat/indic-conformer-600m-multilingual
• Decoder: RNNT (Recurrent Neural Network Transducer)
• Architecture: Conformer (600M parameters)

📝 Notes

• Audio files are automatically resampled to 16kHz mono
• Longer audio files are split into chunks for parallel processing
• GPU acceleration is automatically used when available
• Maximum audio duration is 30 minutes per request

🤝 Credits

Built with:

• AI4Bharat IndicConformer
• Hugging Face Transformers
• FastAPI

📄 License

MIT License