https://github.com/kingabzpro/simple-mlops-with-urdu-asr
A beginner-friendly project for building, testing, and deploying an Urdu ASR (Automatic Speech Recognition) model.
https://github.com/kingabzpro/simple-mlops-with-urdu-asr
Last synced: about 1 month ago
JSON representation
A beginner-friendly project for building, testing, and deploying an Urdu ASR (Automatic Speech Recognition) model.
- Host: GitHub
- URL: https://github.com/kingabzpro/simple-mlops-with-urdu-asr
- Owner: kingabzpro
- License: apache-2.0
- Created: 2025-06-28T12:46:37.000Z (3 months ago)
- Default Branch: main
- Last Pushed: 2025-07-11T13:19:02.000Z (3 months ago)
- Last Synced: 2025-07-11T15:45:09.030Z (3 months ago)
- Language: Jupyter Notebook
- Size: 66.4 KB
- Stars: 1
- Watchers: 0
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# 🎤 Urdu ASR – Modern MLOps Pipeline
[](LICENSE)
[](https://www.python.org/downloads/)
[](https://bentoml.com/)
A production-ready end-to-end MLOps pipeline for Urdu Automatic Speech Recognition using fine-tuned Whisper-v3, featuring:
- 🔄 **Automated Pipeline**: Prefect orchestration for data processing, training, and deployment
- 📊 **Experiment Tracking**: MLflow integration for comprehensive model tracking
- 🚀 **Modern Serving**: BentoML 2.0+ with observability and monitoring
- 🎯 **Interactive UI**: Gradio interface for easy testing and demonstration
- 📈 **Observability**: Built-in metrics, logging, and health monitoring
## ✨ Features
- **Fine-tuned Whisper-large-v3** on Common Voice Urdu dataset
- **Real-time transcription** with audio duration and processing time metrics
- **Comprehensive error handling** with detailed error types and messages
- **Health monitoring** and service status endpoints
- **Scalable deployment** with configurable resources and concurrency
- **Development-friendly** with hot reloading and debugging support
## 🚀 Quick Start
### Prerequisites
```bash
# Python 3.11+
python --version
# Install dependencies
pip install -r requirements.txt
# Set up environment variables
export HF_TOKEN="your_huggingface_token"
export MLFLOW_TRACKING_URI="your_mlflow_uri" # optional
```
### Run the Complete Pipeline
```bash
# Execute the full MLOps pipeline
prefect worker start --pool default-agent-pool &
prefect deploy --name urdu-asr-pipeline
prefect deployment run asr_pipeline/urdu-asr-pipeline
# Or run locally
python prefect/pipeline.py
```
### Serve the Model
```bash
# Build and serve locally
bentoml build
bentoml serve urdu-asr:latest
# Or deploy to BentoCloud
bentoml deploy urdu-asr:latest
```
### Launch Gradio Interface
```bash
# Update endpoint in gradio_app/app.py
# Then launch the interface
python gradio_app/app.py
```
## 📋 Pipeline Steps
1. **Data Download** (`scripts/download_dataset.py`)
- Downloads Common Voice v17 Urdu dataset
- Handles authentication with HuggingFace
2. **Data Preprocessing** (`scripts/preprocess_dataset.py`)
- Filters for valid Urdu text using regex
- Resamples audio to 16kHz mono FLAC
- Parallel processing for efficiency
3. **Model Training** (`scripts/train.py`)
- Fine-tunes Whisper-large-v3 on Urdu data
- MLflow experiment tracking
- Automatic checkpoint management
- Early stopping and best model selection
4. **Model Evaluation** (`scripts/evaluate.py`)
- Calculates Word Error Rate (WER)
- Generates prediction samples
5. **Model Serving** (`bento_service/service.py`)
- BentoML service with observability
- Health checks and monitoring
- Error handling and validation
## 🔧 Configuration
### Environment Variables
```bash
# Required
export HF_TOKEN="your_token"
# Optional - Training
export MODEL_ID="openai/whisper-large-v3"
export NUM_EPOCHS="3"
export TRAIN_BATCH_SIZE="16"
export LEARNING_RATE="1e-5"
# Optional - Serving
export BENTO_ENDPOINT="http://localhost:3000/transcribe"
export MAX_FILE_SIZE_MB="25"
# Optional - MLflow
export MLFLOW_TRACKING_URI="sqlite:///mlflow.db"
```
### Resource Configuration
The BentoML service is configured for:
- **Memory**: 2Gi
- **GPU**: 1 (if available)
- **Concurrency**: 4 requests
- **Timeout**: 30 seconds
## 📊 Monitoring & Observability
The service includes comprehensive monitoring:
- **Request Metrics**: Count, duration, success/error rates
- **Audio Metrics**: File size, duration, format validation
- **Model Metrics**: Processing time, inference performance
- **Health Checks**: Service status and model availability
Access metrics at:
- Health: `GET /health`
- Metrics: `GET /metrics` (Prometheus format)
- Docs: `GET /docs` (OpenAPI)
## 🧪 Testing
```bash
# Test the service locally
curl -X POST "http://localhost:3000/transcribe" \
-F "audio_file=@test_audio.wav"
# Health check
curl "http://localhost:3000/health"
```
## 📁 Project Structure
```
.
├── bento_service/
│ └── service.py # Modern BentoML service
├── gradio_app/
│ ├── app.py # Enhanced Gradio interface
│ └── requirements.txt # UI dependencies
├── scripts/
│ ├── download_dataset.py # Data acquisition
│ ├── preprocess_dataset.py # Data processing
│ ├── train.py # Model training
│ └── evaluate.py # Model evaluation
├── prefect/
│ └── pipeline.py # Orchestration pipeline
├── bentofile.yaml # BentoML configuration
├── requirements.txt # Core dependencies
└── README.md
```