Translation. Region: Russian Federation –
Source: Novosibirsk State University –
An important disclaimer is at the bottom of this article.
Novosibirsk, November 5, 2025: Researchers at the Novosibirsk State University Center for Artificial Intelligence (NSU CAI) have developed a voice assistant for smart home systems that can operate completely autonomously—without an internet connection. The new device, dubbed "Commander," enhances security and privacy, allowing users to control their devices with their voice without fear of hacking or personal data leakage.
"The key feature of the Commander is that no voice travels beyond the device—everything happens locally, without accessing external servers. This significantly reduces the risk of hacking and data leakage," said Evgeny Pavlovsky, a leading researcher at the NSU Center for Information Security. "This solution is, of course, more expensive, but it's essential for those who are particularly concerned about information security and don't want a microphone in the home eavesdropping and sending recordings to the cloud."
The development was completed by Ahsan Shakur, a master's student in the English-language master's program "Artificial Intelligence and Big Data Analytics" at the Faculty of Mechanics and Mathematics at NSU, under the supervision of specialists from the NSU Center for Information Technologies. In its current version, "Commander" can recognize commands such as "turn on the light," "turn off the fan," or "turn on the lamp in three minutes." But this is just the beginning—the system already supports several AI models and, in the future, will be able to analyze user habits, predicting their actions.
"We're testing the system so it can respond not only to direct commands but also to context. For example, if a person says, 'It's dark,' the device will automatically understand that the light needs to be turned on. All the data stays internal—nothing is transmitted externally," explains Evgeny Pavlovsky.
The device's operation is based on three artificial intelligence models. The first converts speech into text, the second understands the meaning of what is said, and the third executes commands. All of these models run locally, without connecting to cloud services. NSU notes that controlling household appliances doesn't require high computing power: the system can even run on a Raspberry Pi minicomputer.
In the future, "Commander" will be able to control any device that supports standard smart home protocols—HTTP, ZigBee, and LoRaWAN. This means users won't need to install additional software or search for special gadgets—the assistant is compatible with most smart plugs, lamps, and household appliances.
Developers believe that in the age of ubiquitous digitalization, security issues are becoming especially pressing. According to Evgeny Pavlovsky, most modern voice assistants, including well-known commercial solutions, continuously record and send data to external servers. This makes them vulnerable to both leaks and hacker attacks.
"Our idea is different: to create a system that 'hears' but doesn't 'transmit.' It understands commands and controls the home, but remains completely under the user's control. This is the right ideology for smart homes, and the market for such solutions is already emerging," the scientist emphasized.
The NSU Center for Information Systems Research team is currently continuing to refine the technology. The next step will be to integrate all three models into a single system capable of learning and adapting to the owner's habits. The researchers are confident that such autonomous solutions will be an important step toward creating safe and truly "smart" homes of the future.
To protect users from internal system errors, Evgeny Pavlovsky and his colleagues intend to use their Kappa framework for managing datasets and artificial intelligence models, which is designed to create a trusted artificial intelligence management environment.
Please note: This information is raw content obtained directly from the source. It represents an accurate account of the source's assertions and does not necessarily reflect the position of MIL-OSI or its clients.