marceltomas/covid-cough-detection

COVID-19 Cough Audio Dataset (Processed, Kaldi-style)

This repository contains a fully processed, Kaldi-style version of the COVID-19 cough audio dataset used in my COVID-19 detection project.
It includes standardized audio files (16 kHz), a precomputed DTW distance lookup table, and several precomputed feature representations (MFCC, MFSC, Mel spectrograms) to support fast experimentation.

This dataset is intended for researchers and practitioners working on audio classification, medical audio analysis, or benchmarking feature extraction pipelines.

The dataset follows a Kaldi-style directory layout, where all metadata and feature descriptors are stored in text-based index files:

data/
│
├── wavs16k/                       # All audio files resampled to 16 kHz
│
├── train/                         # Training split metadata
│   ├── wav.scp                    # mapping: sample_id -> path_to_wav
│   └── text                       # mapping: sample_id -> label (0/1)
│
├── test/                          # Test split metadata
│   ├── wav.scp
│   └── text
│
└── features/
    │
    ├── lookups/                   # Lookup tables for similarity-based methods
    │   └── dtw_lookup.npy
    │
    ├── train/                     # Precomputed features for training samples
    │   ├── mfcc/
    │   ├── mfsc/
    │   └── mel_spectrogram/
    │
    └── test/                      # Precomputed features for test samples
        ├── mfcc/
        ├── mfsc/
        └── mel_spectrogram/

This dataset is built from this Kaggle competition used in one of my university courses, which is a balanced selection of recordings from 3 different databases: Coswara, Cambridge, and Coughvid. Please ensure compliance with the original Kaggle competition’s terms of use when working with this dataset.