This is an on-going project in which I aim to create a program which can accurately classify common urban sounds, and detect idling engines in the hope that it could be employed in tackling offending vehicles as part of the recent government initiatives to impose "no idling" zones.
This project requires Python 3 and the following packages:
pandas
librosa
glob
matplotlib
numpy
scipy
keras
tensorflow
NB The Training Set used is not included in this repository due to the large file size, however it can be downloaded for free from: https://datahack.analyticsvidhya.com/contest/practice-problem-urban-sound-classification/
- The file
Plot_Audio_Data.pycan be used to show a plot of the waveform for sounds in the training set. The plot below shows the waveform for an Idling Engine recording.
Using a sequential model with 2 hidden layers, I was able to achieve a maximum accuracy of 51% in the cross-validation set. However, removing one of the hidden layers gave a much improved accuracy of 71% in the cross-validation set. Moving on from this, I decided to try using a convoluted neural network on a Mel Frequency Spectogram of the audio clips.
By increasing the number of Epochs to 50, I managed to increase the accuracy to 88% in the cross-validation set.
The next step is to try implementing a Convolutional Neural Network in order to hopefully increase the accuracy to ~100%
The dataset contains the following distribution of data:
-
Sample size: 5434
-
10 different labels
-
Number of samples per label (descending order):
'jackhammer': 668
'engine_idling': 624
'siren': 607
'street_music': 600
'drilling': 600
'dog_bark': 600
'children_playing': 600
'air_conditioner': 600
'car_horn': 306
'gun_shot': 230
The plot below shows the performance of the algorithm on the training and validation sets respectively, for varying hyperparameters of batch size and number of epochs.
