How Sber Taught Its Smart Speakers to Generate Smart Home Scenarios by Voice

Sber has taught GigaChat in its smart speakers to create automation scenarios directly from voice commands. Users can now say, "Create a scenario so the lights and heating switch off when I leave the house," and the speaker generates the automation on its own, without a single touch of the screen.

Voice instead of navigation

Until recently, creating a scenario meant opening the app, finding the right devices in a list, linking them with conditions, and saving the rules by hand. The process was tedious: tapping through the screen three times and hunting for filters put off ordinary users. Now a single phrase is enough.

GigaChat analyzes the user's intent, determines which devices are involved, and generates the scenario in seconds. In essence, this resembles commands like "OK Google, create a routine," but Sber took its own approach and did not choose the classic path of fine-tuning on thousands of examples. Instead, the engineers chose in-context learning: information about the user's specific devices is passed directly into GigaChat's context before generation. The model sees the real topology of the home and works with it without prior retraining. This saves on data labeling and speeds up adaptation to new devices: if a user buys a new lamp, they do not have to wait for a model update.

Personalization is the main problem

The main challenge in smart home control is absolute personalization. One user has 30 devices, another has three. Someone calls a light a "lamp," another says "bedroom light," a third says "sun above the bed." Sensors, switches, custom scripts: all of it can be named in completely different ways.

Ordinary LLMs often stumble on this variability: they cycle through devices on blind guesses, mix up rooms, and misread intent. But an error is unacceptable here; this is not a music recommendation gag. If a scenario fires incorrectly, the user may freeze at night because the heating does not turn on. Or the air conditioner may run during the day in an empty apartment, consuming electricity.

Sber's engineers decided not to retrain the model for each user (which is infeasible) but to give it a full "reference book" in the request context. Before calling GigaChat, the backend assembles a description of all of that user's devices: what functions they have, which room they are in, and what names identify them. GigaChat sees this information and can use it safely.

How it works

The pipeline works roughly like this:

• The user says to the speaker: "Create a good night scenario"
• The speaker recognizes the speech and sends the text to the backend
• The backend requests the user's entire device catalog with function descriptions
• The catalog and the request are passed to GigaChat, which generates a YAML description of the scenario
• The scenario engine validates the result, checking that all devices actually exist
• If the check passes, the scenario is saved and becomes active

Validation at the scenario engine level is a safety net. If GigaChat makes a small mistake (for example, mentions a sensor that does not exist or gets the command syntax wrong), the engine catches it and either fixes it or asks the user to clarify. Sber called this component the scenario engine; it works as an error check for every generated rule.

What this means

The smart home becomes more accessible to ordinary people. If the speaker creates scenarios correctly by voice, a newcomer does not need to memorize an interface or read 50 pages of instructions. Just say what you want, and the system does it.

This is an important step toward moving smart homes out of the enthusiast niche and into a mass market, where people value simplicity above all.