When Russian Military commanders started using unencrypted radio communication, we were able to capture this and deploy our AI tools to extract key insights from the data. What we found were indications of a disorganized army, bogged down, and confused.
Read: “As Russia Plots Its Next Move, an AI Listens to the Chatter,” WIRED, April 4, 2022
Troops frequently go out to the battlefield not fully knowing what they might encounter. Even the best armies in the world rarely have all needed capabilities ready and designed for the task at hand. At Primer, we increase operational readiness by building and deploying flexible AI solutions that are optimized for today’s and tomorrow’s defense and intelligence use cases. To help U.S. and strategic partner Services navigate continuously evolving needs, we at Primer believe that the best approach is not a single perfect AI algorithm. The solution is a powerful AI infrastructure that enables subject-matter experts to build, customize, and deploy AI models that help them rapidly process massive volumes of data. When intercepted Russian audio of military commanders discussing tactics recently became available, we saw an opportunity to test our AI platform capabilities.
Today, linguists and analysts spend their days pouring through noisy radio chatter, on the hunt for actionable information. Analysts are at pains to listen to each minute of audio in hopes of finding useful information and for fear of missing things. Intelligence may only find its way to decision-makers hours later, which could be too late to act on it. Using Primer’s platform capabilities, we can transform these workflows and shorten the time from insight to action. With Primer, users can build and run models across radio chatter that detect Russian language and extract mentions of location, persons, certain topics of interest, or military equipment, summarize text into paragraphs or bullets, and identify trends. By instantly surfacing the audio files and information that is likely most relevant, users cut through the noise, focus on the actual analysis, and develop robust near-real time intelligence for decision-makers.
How we use AI to transform Russian military audio into operational intelligence
We started with audio captured from Software Defined Radio (SDR) feeds transmitted over the internet at http://www.websdr.org/ and connected this to our Primer audio ingestion engine which downloaded the communications as wav files. These audio intercepts were then fed into the Primer Platform. First, we used a noise reduction engine and the audio recordings were cleaned to eliminate static or noise, improving sound quality. Cuts containing music or non-Russian speakers were filtered out as they were unlikely to contain our desired targets. These cuts were converted to English text by first using the Russian speech recognition and then Russian to English translation engines within Primer’s platform. Human experts were then able to review the output to understand any areas where the models may have had difficulty.
With our cleaned and prepared dataset in hand, we pulled in our commercially available national security models to get a better understanding of what was going down at the frontlines.
Here are the analytical machine learning capabilities we used to rapidly extract key information from the Russian radio chatter:
- A topic identification model to find and identify the main audio topics so that analysts can navigate precisely to the audio that they are most interested in.
- A dialog summarizer to give analysts a concise description and overview of what was said across hours of captured audio.
- A “call-to-action” model to find and time-stamp orders being issued.
- A weapons classifier to identify mentions of military equipment.
- A Named Entity Recognition (NER) engine to identify, extract, and disambiguate all the people, locations, and organizations that are mentioned in the radio communications.
Analysts can subscribe to this feed of “calls to action” and consume these in real time on the Primer Platform in apps like Primer Command, or integrate the feed into a Common Operating Picture next to their other battlefield data. This radically democratizes access, as analysts can now directly consume and leverage this information, removing the constraint of needing to be a Russian language specialist or listening for hours to multiple radio frequencies. Instead, analysts can conduct deeper analysis based on a near real time call to action feed of Russian military chatter.
One afternoon’s work on the audio files revealed actionable intelligence that Russian forces were reporting heavy fire in their operating area and losing tanks. They also received orders to retreat from their position. But more importantly we were able to build a custom AI pipeline on our platform to continually and automatically monitor Russian radio chatter and extract tactical insights in near real time.
This type of real-time intelligence can be a game changer for analysts and commanders. Whoever succeeds in deploying operational AI – reliable AI models that are trained, validated, and in production on live operational data during missions – will have the advantage of faster AI-augmented human decision-making, thereby improving the odds at outmaneuvering the adversary.
When it comes to battlefield capabilities, the ability to rapidly customize models that meet constantly evolving needs is a crucial differentiator. The models used above are not perfect and will benefit from further training and specialization for specific national security use cases and niches. Here are a few of the challenges our platform of pre-trained models, labeling, and model training can address to help analysts make immediate workflow improvements in the course of their mission:
- Our dialogue summarizer model had to contend with the idiosyncrasies of tactical radio communications such as speakers repeatedly saying their own name. In the Primer Platform, this model can be retrained to improve performance, by starting with the machine-generated summary a human analyst can correct and improve. This corrected summary turns into another training datapoint to teach and fine tune the model.
- Our weapons classifier model was trained last year during the crisis in the Donbass region of Ukraine. Since then, many more types of military equipment have been deployed. Additionally dialog on radio chatter is much more informal and references to “Helis” are not captured by the initial weapons classifier as “Helicopters.” Users and analyst teams can take the initial models and quickly label new data using our LightTag software and then retrain them to improve their performance in the Primer Platform.
- The geolocation engine was not initially trained on tactical radio communication data. As such it was not initially trained to pick up references to map coordinates which are common on radio chatter for artillery units. The users are able to improve the geolocation model by retraining our Named Entity Recognition engine in LightTag and the Platform to identify these and other geographic references.
How strategic advantage depends on flexible and customizable AI solutions
Deployment of AI in a conflict environment is not about having a single perfect algorithm. Instead, the best solution is often a collection of AI algorithms working together as a complex system — the optimal models will change as quickly as the crisis evolves. The most successful approach is to have flexible and customizable AI tools and infrastructure available to users closest to the need so that they can quickly test, customize, and connect the most performant models together.
At Primer, our goal is to empower those closest to the problem to leverage advanced AI to help them make sense of an incredibly complex information landscape. Machine learning is a force multiplier for linguists and analysts tasked with processing hundreds of hours of audio files. Small teams can process many hours or days worth of intercepts in a matter of minutes. This enables faster and better tactical responses to unfolding events. Intelligence information previously only available hours or days later can now be in the hands of battlefield commanders when it’s needed most. A strong AI infrastructure for Defense must let users retrain ML models quickly – as these opportunities may not last very long.
The ability to rapidly exploit OSINT, even when messy and in other languages, makes it possible for poorly equipped armies to outmaneuver much better trained and resourced ones. That is a huge paradigm shift, the likes of which we have yet come to terms with.
Despite having the best trained and resourced services in the world, the hard truth is that a lack of agility and urgency in integrating and operationalizing AI by the U.S. and our strategic allies means that we fall behind faster than we’re willing to admit. As we look ahead, we’re now in a new era of warfare where those with sophisticated and customizable AI capabilities have a distinct advantage in decision superiority, and thereby the ability to outcompete and outmaneuver adversaries to increase the odds of winning wars that cannot be prevented.
For more information about Primer and to access product demos, contact Primer here.