Sound event detection system

Introduction

This repository contains the code for training and evaluating sound event detection system.
System is trained on the DCASE 2019 Task 4 dataset.

Sound event detection model is trained using PyTorch and PyTorch Lightning frameworks. All the experiments are logged using DVC for reproducibility and artifacts tracking.

Usage

First, download the dataset from the official website and extract it to the DESED_dataset directory. You will have to change the paths in params.yaml to match your directory structure. You can use this command:

sed -i 's|/mnt/d/DESED_dataset|<YOUR_DATA_DIR>/DESED_dataset|g' params.yaml

Directory tree should look like this:

DESED_dataset
├── audio (2)
│   ├── train (4)
│   │   ├── strong_label_real (3470)
│   │   ├── unlabel_in_domain (14412)
│   │   └── weak (1578)
│   └── validation (2)
│       └── validation (1168)
├── dcase_synth (2)
│   ├── audio (2)
│   │   ├── train (1)
│   │   │   └── synthetic21_train (1)
│   │   │       └── soundscapes (30000)
│   │   └── validation
│   │       └── synthetic21_validation (1)
│   │           └── soundscapes (7500)
│   └── metadata (2)
│       ├── train (1)
│       │   └── synthetic21_train (1)
│       └── validation (1)
│           └── synthetic21_validation (2)
└── metadata (2)
    ├── train (4)
    │   └── synthetic21_train (1)
    └── validation (4)
        └── synthetic21_validation (2)

Next, create conda environment with python 3.8.5 and install the required packages using the following command:

conda env create -n env_name python=3.8.5
pip install -r requirements.txt

To train the model, run the following command:

python run_training.py

For model output visualization and error analysis purpose, streamlit app is used. You can run the app using the following command:

cd app/
streamlit run app.py