Yandex Solved the Problem of 'Catastrophic Forgetting' in AI Models — Smart Devices Will Stop Forgetting Commands

Yandex researchers have developed a method to add new voice commands to smart devices without degrading the recognition accuracy of already familiar phrases. The technology is intended for commands that speakers, household appliances, and car assistants execute without an internet connection. Manufacturers of smart devices must regularly expand the set of offline commands, but after training the model on new commands, it may start to perform worse on already familiar queries — this effect is called "catastrophic forgetting." To mitigate it, continuous learning methods are used. Yandex researchers proposed adding a separate small module to the main neural network, which handles new commands, while the model itself remains unchanged and continues to recognize previously familiar phrases with the same accuracy. The new approach was tested on the Google Speech Commands voice command set: the average miss rate for new phrases decreased by nearly a third — from 6.46% to 4.37%, while the model size increase required for the method was less than 10%. The study was accepted for the Interspeech 2026 conference, which will be held from September 27 to October 1 in Sydney.