Yandex trains its smart speaker new fast commands without losing quality of old ones

Yandex’s smart speaker (e.g., Station Street) supports fast commands — short phrases like 'next' or 'quieter' — that bypass the activation word 'Alice' and run locally on the device. To add new commands without harming old ones, the team rejected full retraining (risk of regression), separate small models (inefficient and redundant), and classic continual learning methods like EWC (no guarantees). Instead, they used a modular network expansion: starting from a certain layer of the frozen original model, a parallel branch extracts activations, computes a small additive vector, and concatenates it at each step. This allows learning new commands while preserving old ones, with full control over memory and compute. Training used two stages — pretraining on 60–100 million uncurated ASR logs and fine-tuning on 400,000–1 million clean in-home recordings. Results: CPU load reduced by 35%, RAM by 50%, and false activations dropped by 60%. The work was published at Interspeech 2026.