A high-performance FastAPI backend service for Persian speech recognition using the jonatasgrosman/wav2vec2-large-xlsr-53-persian model.

• 🚀 FastAPI: Modern, fast web framework with automatic API documentation
• 🇮🇷 Persian Language Support: Specialized for Persian speech recognition
• 🎯 High Accuracy: 30.12% WER and 7.37% CER on Common Voice Persian test set
• 📱 Mobile Ready: Optimized for React Native integration
• 🔄 Async Support: Full async/await support for better performance
• 📊 Confidence Scoring: Provides transcription confidence levels
• 🎵 Multiple Formats: Supports WAV, MP3, M4A, FLAC, OGG, AAC
• 📖 Auto Documentation: Interactive API docs at /docs
• 🔧 GPU Support: Automatic GPU detection and utilization

1. Install Python 3.10+

   python --version  # Should be 3.10 or higher

2. Create virtual environment (recommended)

   python -m venv venv
   source venv/bin/activate  # On Windows: venv\Scripts\activate

3. Install dependencies

   pip install -r requirements.txt

1. Start the server

   python main.py
   # or
   python run.py

2. Access the API

   • Server: http://localhost:8000
   • Interactive docs: http://localhost:8000/docs
   • Alternative docs: http://localhost:8000/redoc

• GET /health
• Returns service status and model information

• POST /transcribe
• Upload audio file via multipart/form-data
• Content-Type: multipart/form-data
• Parameter: audio (file upload)

• POST /transcribe-base64
• Send base64 encoded audio data
• Content-Type: application/json
• Body: {"audio_base64": "base64_encoded_audio"}

• POST /transcribe-url
• Transcribe audio from a URL (for testing)
• Content-Type: application/json
• Body: {"audio_url": "https://example.com/audio.wav"}

{
  "success": true,
  "transcription": "متن تبدیل شده به فارسی",
  "confidence": 0.95,
  "language": "persian",
  "model": "jonatasgrosman/wav2vec2-large-xlsr-53-persian",
  "audio_duration": 5.2,
  "processing_time": 1.8
}

• Sample Rate: Automatically converted to 16kHz
• Duration: 0.5 - 30 seconds
• Formats: WAV, MP3, M4A, FLAC, OGG, AAC
• Quality: Higher quality audio produces better results

• First Request: ~5-10 seconds (model loading)
• Subsequent Requests: ~1-3 seconds
• GPU Acceleration: Automatic if CUDA is available
• Concurrent Requests: Supports multiple simultaneous requests

Run the test suite:

python test_api.py

./deploy_to_server.sh

This will:

1. Install Python 3.10 and dependencies
2. Create virtual environment
3. Install FastAPI dependencies
4. Set up systemd service
5. Start the service on port 8001

For aiapp.sazjoo.com domain:

# Copy nginx config to server
scp nginx-aiapp-config.conf [email protected]:/etc/nginx/sites-available/aiapp.sazjoo.com

# On server:
ln -s /etc/nginx/sites-available/aiapp.sazjoo.com /etc/nginx/sites-enabled/
nginx -t
systemctl reload nginx

Update your React Native app's backend URL to:

http://aiapp.sazjoo.com

Example usage in React Native:

const response = await fetch('http://aiapp.sazjoo.com/transcribe-base64', {
  method: 'POST',
  headers: {
    'Content-Type': 'application/json',
  },
  body: JSON.stringify({
    audio_base64: base64AudioData,
  }),
});

const result = await response.json();
console.log('Transcription:', result.transcription);

• Ensure stable internet connection for initial model download
• Check available disk space (model is ~1.5GB)
• Verify Python version compatibility

• Check audio file format is supported
• Ensure audio duration is within limits (0.5-30 seconds)
• Verify audio file is not corrupted

• Install CUDA for GPU acceleration
• Increase server resources for better performance
• Consider using multiple workers in production

• Model: jonatasgrosman/wav2vec2-large-xlsr-53-persian
• Base Model: Facebook's wav2vec2-large-xlsr-53
• Training Data: Common Voice 6.1 Persian dataset
• Performance: 30.12% WER, 7.37% CER on test set
• Input Requirements: 16kHz audio
• Output: Persian text transcription

This service uses the wav2vec2 model which is subject to its own license terms.