Sound20

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.

Data

You can find spectrograms of samples in spectrogram_data.

Statistics

Split name	Train set	Validation set	Test set
Number of samples	16,636	3,249	3,727

Labels

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

Label	Description
0	Drum_FloorTom
1	Drum_HiHat
2	Drum_Kick
3	Drum_MidTom
4	Drum_Ride
5	Drum_Rim
6	Drum_SmallTom
7	Drum_Snare
8	Guitar_3rd_Fret
9	Guitar_9th_Fret
10	Guitar_Chord1
11	Guitar_Chord2
12	Guitar_7th_Fret
13	Bufo_Alvarius (a type of toads)
14	Bufo_Canorus (a type of toads)
15	Pseudacris_Crucifer (a type of frogs)
16	Allonemobius_Allardi (a type of crickets)
17	Anaxipha_Exigua (a type of crickets)
18	Amblycorypha_Carinata (a type of katydid)
19	Belocephalus_Sabalis (a type of katydid)

Usage

Loading data

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')

Evaluate using CNNs

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.

Experimental results

The experimental results using LeNet and VGG_F network structure.

Network structure	Testing Accuracy
LeNet	78.07%
VGG_F	79.15%

References

Recognizing Sounds (A Deep Learning Case Study)

Animal Sound Data