WhisperX API Server

A lightweight Flask API server that wraps WhisperX for efficient audio transcription using GPU acceleration. This service maintains a loaded model in memory for faster processing and provides a simple HTTP interface for transcription requests.

Context

WhisperX has very specific dependencies which run on older python and torch versions, and is usually incompatible with some of the newer applications, so I separated out WhisperX, and made it a simple Flask API server. That way my code in other applications can just send API request to this server, and receive transcription as response.

Features

• Persistent Model Loading: Keeps WhisperX model loaded in memory for faster processing
• GPU Acceleration: Utilizes CUDA for efficient transcription
• RESTful API: Simple HTTP interface for transcription requests
• Health Monitoring: Endpoint to check service status and GPU memory usage

Prerequisites

• Docker
• NVIDIA GPU with CUDA support
• NVIDIA Container Toolkit
• Audio files accessible from the host machine

Quick Start

1. Clone the repository

git clone https://github.com/yourusername/whisperx-api-server.git
cd whisperx-api-server

2. Build the Docker image

docker build -t whisperx-api .

3. Run the container

docker run --gpus all --shm-size=8gb --runtime=nvidia --network=host --volume="/sizzle_storage:/sizzle_storage:rw" --volume="$(pwd):/app" -p 5000:5000 --name whisperx-api whisperx-api

4. **Wait After running the container, it takes about 5 minutes while for Whisper model to download and load. Wait till these messages appears in screen:

 * Serving Flask app 'app'
 * Debug mode: off
INFO:werkzeug:WARNING: This is a development server. Do not use it in a production deployment. Use a production WSGI server instead.
 * Running on all addresses (0.0.0.0)
 * Running on http://127.0.0.1:5000
 * Running on http://192.168.3.193:5000
INFO:werkzeug:Press CTRL+C to quit

Once the above message appears, you are good to start making requests for API calls.

You will also be able to see a process called WhisperX-API loaded in GPU memory once you run the command nvidia-smi.

API Endpoints

POST /transcribe

{
    "file_path": "/audio/your-audio-file.wav"
}

Health Check

GET /health

{
    "status": "healthy",
    "model_loaded": true,
    "gpu_memory_allocated": 4521.5,
    "gpu_memory_cached": 6144.0
}

Usage Examples using Python

import requests

def transcribe_audio(file_path):
    url = "http://localhost:5000/transcribe"
    payload = {"file_path": file_path}
    response = requests.post(url, json=payload)
    return response.json()

# Example usage
result = transcribe_audio("/audio/sample.wav")
print(result["segments"])

Directory structure

whisperx-api-server/
├── Dockerfile
├── README.md
├── app.py

Bibliography

Whisper

Flask