Yandex researchers reveal how neural networks recognize voice commands in noisy environments

Yandex researchers have released a scientific paper detailing a neural network technology that can recognize voice commands even in noisy environments. Already deployed in Yandex smart devices, the technology's key principles are now available to developers worldwide. The paper has been accepted for presentation at Interspeech 2025 — a leading global conference on spoken language processing and speech technology — which will be held from August 17 to 21, 2025, in Rotterdam, the Netherlands. Microsoft, Google DeepMind, Google AR, and other prominent tech companies and research institutions will also be presenting at the event. 

The technology detailed in the paper enables Yandex smart devices to accurately detect voice commands amidst a wide range of background noises, including music, running water, social gatherings, and construction work outside. Users can issue commands without the need to pause appliances like vacuum cleaners or raise their voices over music. This innovation will enable companies worldwide to accelerate the development of voice assistants and voice-controlled devices, thereby reducing false activations and enhancing user experience through easier voice control.

Conventional smart devices and AI assistants rely on echo suppression algorithms to filter background music and noise reduction algorithms to minimize background sounds. However, noise reduction can compromise speech clarity. To solve this problem, Yandex developed a neural network attention mechanism that simultaneously receives two input signals — one with noise reduction and one with echo suppression. The neural network then selects the clearest signal to ensure reliable voice recognition across a wide range of background noises.

Dmitry Solodukha,
Head of Voice Recognition at Yandex

Until now, there's been no standard approach to voice recognition in noisy environments that performs reliably both in lab tests and in real-world conditions. Many companies and researchers face similar challenges, but lack access to commercial solutions, which forces them to spend resources building such solutions from scratch. By publishing our method, we hope to accelerate innovation in voice interfaces, help others avoid common mistakes, and ultimately enable the development of more convenient and reliable voice–controlled devices.