This is a dataset including 20 animal and instrument sounds. This dataset is constructed using Animal Sound Data and Instrument Data. Each audio is split into multiple samples, and we make sure that samples in Train, Validation, Test sets are disjoint and separated.

You can find spectrograms of samples in spectrogram_data.

Each sample is assigned to one of the 20 labels, which include sounds of drums, guitars, frogs, and insects.

You can load this dataset using Python with Numpy.

import numpy as np 
x_train = np.load('spectrogram_data/train_X.npy')
y_train = np.load('spectrogram_data/train_Y.npy')

We conduct two experiments on this dataset using LeNet and VGG_F network structures. To run the experiments, please using the following commands.

For LeNet, use

sh scripts/run_LeNet.sh 

For VGG_F, use

sh scripts/run_VFF_F.sh 

Note that each script will run training and testing procedures and store the Train, Val, Test accuracy.

The experimental results using LeNet and VGG_F network structure.

Recognizing Sounds (A Deep Learning Case Study)

Animal Sound Data