Using AI and ML models, we predict all sorts of sounds of war with only the sound signature to go on. With a lightweight model that can either run in the cloud or directly embedded on edge devices with a microphone, Sounds of War models are able to detect a multitude of sounds from fairly poor quality faint signals.

Sounds of War uses sound rather than computer vision to make its inferences. This allows it to detect events on a battle space without having direct line-of-sight, enabling deployed prediction devices to be hidden and/or buried in the background.

Mission

To provide powers and individuals fighting for justice with additional high-tech situational-awareness tools in their ongoing battles.

Each prediction provides crucial early-warning data helpful when planning and issuing orders to general populations, armies, and armed groups.

The dataset

The current model is trained on over 200 gigabytes of sound data from various modern conflicts around the world, with a significant portion coming from crowd-sourced data recorded during the war in Ukraine.

The data continues to be scraped from various social media sources, including Twitter, Reddit, Telegram, and others. Some data sent directly to Sounds of War is also used during the training.

The model

The model is based on a pre-trained neural network finely tuned on a large dataset of sound to detect specific sounds of war. The model is currently able to detect some categories of sounds with up to 95% accuracy, with an average of 84% across all categories.

output
Waveforms of explosions, AK47, and other gunfire sounds.
output
Spectograms of different explosions and gunfire sounds.

The model can currently detect and classify multiple sounds, including:

  • Vehicles
    • Tanks (including some specific types)
    • Armored personnel carriers
    • Artillery
    • Anti-aircraft
    • Missiles
    • Jets
    • Helicopters
    • Other heavy vehicles
  • Gunfire
    • AK47/AK74
    • Glock 45
    • Ruger
    • FSD50k
    • Other arms
  • Rockets
    • Missiles
    • RPGs
    • Javelins
    • MANPADs
  • Individuals conversing
  • Radio chatter
  • City-wide sirens

Get and deploy the model

Due to the potential adversarial use-cases, the model is available only request. Please contact us using the contact form below if interested.

The model has been designed in a way that allows deployment on edge devices, so that inference can be made offline at any time in remote locations. These predictions can then be sent to central command, which can then process and act on the incoming information.

Contribute

Sounds of War is a project that is built and run by volunteers. If you are interested in supporting the project, we're constantly looking for people to help out, with a focus on the following tasks:

  • Data de-duplication
  • Data labelling

If you can help out in any other way too, we're interested! If you've got marketing or business skills, are knowledgeable about Tensorflow and Machine Learning, or have connections that can deploy the Sounds of War model in real-world environments - please reach out.

Contact

If you would like access to the model or would like to contribute to the Sounds of War project, please contact us.