Refer to the supporting webpage here: https://apps.ahlab.org/DroneAudioSet-code/

Dataset available here: https://huggingface.co/datasets/ahlab-drone-project/DroneAudioSet

Use the pandas or datasets libraries to download the dataset from HuggingFace.

The fully downloaded dataset, according to the path variable, would look like this:

ComputeResourcesCheck folder contains all sample audio files

• ComputeResourcesCheck/preprocessed-audio: contains 6 audio file recordings containing source and drone sounds. The chosen setting was:

Volume: 80pc
Room: room1
Drone: drone1
Drone-Speaker Distance: speaker-dist-1m
Mic: mic3_8array-up
Drone-Mic Distance: mic-dist-25cm

• ComputeResourcesCheck/beamforming: contains 6 audio files after the preprocess-audio files are passed through the beamforming (MVDR) stage.
• ComputeResourcesCheck/spectral-gating: contains 6 audio files after the beamforming files are passed through the spectral gating (noise-reduce) stage.
• ComputeResourcesCheck/mpsenet: contains 6 audio files after the beamforming files are passed through the neural noise suppression (MPSENET) stage.
• ComputeResourcesCheck/classification: contains 6 audio files after the mpsenet files are passed through the classification (SSLAM) stage.

Go to scripts/

conda create -n droneaudioset python=3.9 conda activate droneaudioset pip install -r requirements.txt

For mpsenet and sslam, separate environments are required, as provided in requirements_mpsenet.txt and requirements_sslam.txt. For more details, refer to:

MPSENet Github: https://github.com/yxlu-0102/MP-SENet/tree/main

SSLAM Github: https://github.com/ta012/SSLAM/tree/main

• create a sub-folder mpsenet in the scripts folder
• copy the folder models, and the python files dataset.py, env.py, inference.py, and utils.py from MPSENET repo (https://github.com/yxlu-0102/MP-SENet/tree/main) into the scripts/mpsenet folder
• also copy over best_ckpt from MPSENET repo to scripts folder

• create a sub-folder SSLAM in the scripts folder
• copy the folder SSLAM_inference and the files checkpoint_best.pt and label_descriptors.csv from SSLAM repo (https://github.com/ta012/SSLAM/tree/main) to scripts/SSLAM folder

• audioset_labelmapping.csv: has the mapping of the 527 audioset classes to the three categories HV (human vocals), HNV (human non-vocals), and NH (non-human) sounds
• avg-snr-per-setting.csv: has the average SNR computed per setting
• docs: contains all files for the webpage