AI-based system for real-time detection of whip sounds in horse racing

© Aaron Johnson Dreamstime.com

  

Researchers at the University of Tsukuba have developed an artificial intelligence–based system capable of
automatically detecting whip sounds in horse racing, offering a potential alternative to the labour-intensive manual review processes currently used to enforce whip regulations.

In many racing jurisdictions, the use of whips is strictly regulated to protect animal welfare and ensure fair competition. Violations, such as excessive force or exceeding the allowed number of strikes, are typically identified by race stewards through careful frame-by-frame analysis of video footage. While effective, this approach is time-consuming, costly, and impractical for real-time enforcement during live races. The new study addresses these limitations by focusing on the acoustic signature of whip strikes rather than visual evidence.

Whip sounds are highly impulsive and contain very high-frequency components that are difficult to capture using conventional audio recording systems. To overcome this challenge, the research team recorded race audio at an unusually high sampling rate of 192 kHz, enabling precise capture of both the high-frequency content and the fine temporal structure of whip strikes. Audio data were collected from 24 official horse races held in Japan, yielding a dataset that included 620 carefully annotated whip strike events.

Using this dataset, the researchers built an automated system to detect whip sounds in race audio. The system was trained to recognise the acoustic patterns of whip strikes and how these sounds change over time. Several model designs were tested to find the most effective way to detect the very short, high-pitched nature of whip sounds.

One major challenge was that whip strikes occurred far less often than background noise such as crowd sounds and hoofbeats. To prevent the system from being biased toward this background noise, the researchers reduced the amount of non-whip audio used during training. The best-performing model correctly identified whip strikes with an accuracy score of 69.8%.

Beyond accuracy, the study also examined processing speed. Offline evaluations revealed that the best-performing model could analyse audio faster than real time under many conditions, suggesting that live race monitoring is technically feasible. 

The research provides the first clear confirmation that whip sounds contain critical very high-frequency elements, underscoring the importance of high-sampling-rate audio for this application. At the same time, the authors acknowledge remaining challenges, including environmental noise in race settings and the relatively small size of the current dataset, which can affect robustness and generalization.

Overall, the study establishes an approach for automatic whip strike detection using sound event detection and deep learning. With further data collection and improvements in noise robustness, the system could support real-time rule enforcement, promote fairer competition, and contribute to improved animal welfare in horse racing.

 

For more details, see:

Aoi Taguchi, Yuki Fujita, Keiichi Zempo,

Whip strike detection using high-sampling-rate audio by evaluating convolutional recurrent neural network configurations and class imbalance strategies,

Engineering Applications of Artificial Intelligence (2026) Vol 164, Part A,113272,

ISSN 0952-1976,

https://doi.org/10.1016/j.engappai.2025.113272