Cyrb sound recognition

Municipality of The Hague

Algorithm that detects a sound and determines what kind of sound it is. The type of sound is recorded.

Last change on 23rd of August 2024, at 16:03 (CET) | Publication Standard 1.0

• Publication category: Impactful algorithms
• Impact assessment: Field not filled in.
• Status: Out of use

General information

Theme

Public Order and Safety

Begin date

Field not filled in.

Contact information

[e-mail]

Responsible use

Goal and impact

This is a pilot project. The algorithm was developed to recognise sounds. In addition, it aims to get a picture of where, when and what types of noise are present. A third aim is to possibly act immediately if the type of noise or noise level gives cause to do so.

The results of the algorithm have been validated with residents to investigate whether the perception of noise matches the measured values. If a citizen/visitor/business causes a noise nuisance or a disturbing type of noise (e.g. gunshot), the citizen will have to deal with enforcement or police.

Considerations

There is no alternative; this is the simplest and least intrusive measurement.

Human intervention

Yes. The sensor can be turned off in a simple way. This will be decided if there is no added value or if objections arise that need to be taken into account.

Risk management

The algorithm is trained on public datasets of similar sounds. Because these do not always sound the same and have the same characteristics (and nature) as the sounds captured on location, the algorithm could make errors in sound recognition. As the algorithm is for imaging purposes only, this does not pose any immediate risks for the citizens and entrepreneurs involved. There is no risk of privacy violation because the sound itself is not stored. The categorisation of the sound does not include aspects that could lead to bias, e.g. nationality.

Legal basis

Municipal Act. Article 172 Enforcement of public order

Operations

Data

Sound clips

Technical design

In evaluating the model, it uses the confusion matrix and the categorical-crossentropy as a metric.

The algorithm converts 4-second sound clips into spectrograms, visual representations of the sound. The spectrograms are images of 128x128 pixels in greyscale. These spectrograms are then classified by an image recognition algorithm into a number of categories, such as "stationary traffic", "music", "honking", etc. This image recognition model is a relatively small convolutional neural network of five layers. (Three convolutional layers followed by max-pooling and two fully connected layers). The model runs in python, using the Tensorflow library.