AudioLens

This project is an end-to-end audio classification system built with a deep ResNet-style CNN that converts audio into Mel Spectrograms for image-based sound recognition. It features advanced training techniques like Mixup augmentation, Frequency Masking, AdamW optimization, and OneCycleLR scheduling, all backed by TensorBoard for performance monitoring. The trained model is deployed as a serverless GPU-powered FastAPI endpoint via Modal with a sleek Next.js dashboard built using Tailwind CSS and Shadcn UI offering real-time classification, confidence scores, waveform and spectrogram visualization, and CNN feature map inspection.

Setup

Follow these steps to install and set up the project.

Install Python

Download and install Python if not already installed. Use the link below for guidance on installation: Python Download

Create a virtual environment with Python 3.12.

Backend

Navigate to folder:

cd audio-cnn

Install dependencies:

pip install -r requirements.txt

Modal setup:

modal setup

Run on Modal:

modal run main.py

Deploy backend:

modal deploy main.py

Frontend

Install dependencies:

cd audio-cnn-visualisation
npm i

Run:

npm run dev