Sber is developing a new type of artificial intelligence technology – “full-duplex” – which allows AI to simultaneously listen to the interlocutor and speak, just like a human does during a phone conversation. Sergey Markov, Director for AI Technology Development at Sberbank, spoke about the new development.

Today, most voice assistants operate on a queue principle. A person first speaks, then the system processes the request, and only after that does it respond. With the new technology, they are trying to arrange this process differently.

The main difference of full-duplex is that instead of several separate neural networks, one model is used. Currently, one system translates voice into text, another forms a response, and a third converts it back into speech. This allows AI to simultaneously listen and respond.

If video modality is added, such a system will also be able to see the interlocutor and control the avatar's movements.

Similar technologies have already appeared abroad. The first was the Moshi model from the French startup Kyutai in 2024. Then Google presented its solutions with Gemini Live and OpenAI with GPT-Live.

Full-duplex significantly reduces response latency – from several seconds to 160–320 milliseconds. This is already comparable to the reaction speed of a human in conversation.

In Russia, no one has previously publicly announced a similar development. Yandex has a close technology, Realtime API, with a fast response and the ability to interrupt the assistant, but, according to Markov, this is not a full-fledged full-duplex.

The difference is that the Yandex system waits for a change in the replica, while a full-fledged full-duplex continues to listen to the user even when it is already speaking itself.