In this technical report, we describe our submission for Task 1 Data-Efficient Low-Complexity Acoustic Scene Classification [1]. We adopt the baseline model and add a specialised knowledge distillation process before proceeding with the baseline training process. Our model was distilled on a pruned subset of the AudioSet dataset using large pretrained models. The pruning of the dataset is based on the similarity of the data to the targeted challenge dataset.
Zitation:
Werning, A., & Haeb-Umbach, R. (2023). UPB-NT submission to DCASE24: Dataset pruning for targeted knowledge distillation [Technical report]. Detection and Classification of Acoustic Scenes and Events (DCASE) Challenge 2024. https://dcase.community/documents/challenge2024/technical_reports/DCASE2024_Werning_48_t1.pdf