Sound Classification using CNN

Sample code to show how to use the convolutional Neural Network (CNN) to build a model to classify sound

Sound Dataset

UrbanSound8K https://urbansounddataset.weebly.com/urbansound8k.html

Remark: There are 2 set of sounds which are UrbanSound and UrbanSound8K. Please download UrbanSound8K

Prerequisites

• Python 2.7
• Keras
• Librosa
• Numpy
• Pandas
• Virtualenv

IDE

VsCode https://code.visualstudio.com/download

Setup Environment

I have only run the program on Mac OS

1. Use git or download the source code
2. Setup virtualenv in the source root folder by running virtualenv venv
3. Run source venv/bin/activate
4. Run pip install -r requirements.txt to install the library or pip install {library} one by one
5. You can run python predictUrbanSound8K.py to test whether the environment is correctly setup. The program will try to predict the wav files inside predict folder and returns predict result.
6. Download SoundDataSet from UrbanSound8K: https://urbansounddataset.weebly.com/urbansound8k.html
7. Move UrbanSound8K to the source root folder

How to train and predict using UrbanSound8K

Train Model Please run python trainUrbanSound8K.py The program will try to exract all wav features from UrbanSound8K folder, then build the model using CNN. At the end the program will export the models to folder in .h5 extension. You can use the model to predict .wav files.

Predict Please rename the model you built to urban-sound.h5 and you can use it to predict .wav files.

First put the .wav files into predict folder. Then run python predictUrbanSound8K.py

The result will show confidence level for each labels and the predict result e.g.

air_conditioner 0.00%
car_horn 0.07%
children_playing 0.00%
dog_bark 0.00%
drilling 99.92%
engine_idling 0.00%
gun_shot 0.00%
jackhammer 0.00%
siren 0.00%
street_music 0.01%
Result: drilling

Program Logic

The program logic is straightforward. It use librosa (the sound library) to convert the sound to spectrogram. Then it use the spectrogram as the sound feature.

Then it will reshape the data in 128 x 128 and use as an input format for the conventional Neural Network (CNN) to build the model to classify sound.

For each wav files, the naming convension must be in {wav-name}-{label}, because the program will use string split('-') to get the class label from the wav file. The class label will be used to teach the CNN the correct category for the sound file.

Build Your Own Sound Classification

You can create your own dataset and import the data to the program like UrbanSound8K.

You can use train.py and predict.py as an reference. Also you need to update the class.csv for the sound classiifcation list.