BelSU Researchers Develop AI That Hears Danger in Urban Noise

Researchers at Belgorod National Research University (BelSU) have developed a software module capable of instantly recognizing the sounds of gunshots and breaking glass, even in noisy environments where the human ear struggles. The system called Automated System for Monitoring Alarm Situations uses convolutional neural networks to analyze audio spectrograms, essentially teaching artificial intelligence to read danger signs from sound patterns.

The project emerged from the national Priority-2030 Programme building on existing research into hybrid speech signal analysis.

“We’re implementing a project on recognizing emotions in speech. When the project on recognizing aggression from speech signals emerged, I suggested the students develop something relevant to security – recognizing the sounds of breaking glass and gunshots,” explains Tatyana Balabanova, head of the BelSU Department of Automated Systems and Technologies.

The development began as a graduation project by Anastasia Kuzichkina and is now continued by third-year student Georgy Kantsiber. By the end of 2026, the team plans to test the system in a real urban environment to confirm its effectiveness outside laboratory conditions.

What sets this software apart is its resistance to interference. The neural network was trained on over 250 natural noises, including roosters crowing, cars humming, and human voices. This training makes it resistant to acoustic illusions that can fool classic frequency analysis systems. 

“Our program doesn’t confuse the sound of breaking glass with other sounds, even if they are similar. Some passing-by cars make sounds very similar to gunshots,” Balabanova emphasizes.

Unlike the American ShotSpotter system, whose operating principles remain undisclosed, BelSU’s solution is a purely software-based approach using lightweight convolutional neural networks. The entire program takes up only about 200 MB, meaning it can run on a standard personal computer with a sound card – no supercomputers or specialized hardware required.

This compact design enables another strategic advantage: integration into existing infrastructure. Most modern surveillance cameras already have built-in microphones. The program can detect dangerous activity using sound even where cameras cannot see – in blind spots, around corners, or deep within crowds.

The team’s ambitions extend further. They plan to add detection of aggressive speech and shouting, recognizing that verbal escalation almost always precedes physical danger. Such capabilities could make the system valuable for airports, train stations, educational institutions, and large public events.