2017/12/27 by Shao-Yen Tseng, Tseng, Shao-Yen, Juncheng Li +9
Computer Science · #Audio and Speech Processing (eess.AS) #FOS: Computer and information sciences #FOS: Electrical engineering #Music and Audio Processing #Sound (cs.SD) #Speech Recognition and Synthesis #Speech and Audio Processing #electronic engineering #information engineering
paper · pdf · doi:10.48550/arxiv.1712.09673
openalex publication_date 2017/12/27 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
State-of-the-art audio event detection (AED) systems rely on supervised\nlearning using strongly labeled data. However, this dependence severely limits\nscalability to large-scale datasets where fine resolution annotations are too\nexpensive to obtain. In this paper, we propose a small-footprint multiple\ninstance learning (MIL) framework for multi-class AED using weakly annotated\nlabels. The proposed MIL framework uses audio embeddings extracted from a\npre-trained convolutional neural network as input features. We show that by\nusing audio embeddings the MIL framework can be implemented using a simple DNN\nwith performance comparable to recurrent neural networks.\n We evaluate our approach by training an audio tagging system using a subset\nof AudioSet, which is a large collection of weakly labeled YouTube video\nexcerpts. Combined with a late-fusion approach, we improve the F1 score of a\nbaseline audio tagging system by 17%. We show that audio embeddings extracted\nby the convolutional neural networks significantly boost the performance of all\nMIL models. This framework reduces the model complexity of the AED system and\nis suitable for applications where computational resources are limited.\n