This thesis investigates using knowledge distillation (KD) to improve sound event detection model performance. Sound event detection is challenging due to complex audio signals, limited data, and a large number of sound classes. Traditional approaches involve large, complex models, which require significant computational resources. KD can train smaller, more efficient models to mimic the behavior of larger models, allowing for improved performance with fewer resources.
Алхаддур А. (науч. рук. Матвеев Ю.Н.) Алгоритмы уменьшения размерностей моделей для анализа звуковых событий методом дистилляции знаний // Сборник тезисов докладов конгресса молодых ученых. Электронное издание. – СПб: Университет ИТМО, [2023]. URL: https://kmu.itmo.ru/digests/article/11597