Weakly labeled acoustic event detection using local detector and global classifier

Acoustic event detection plays an important role in the description of acoustic contents and the computational auditory scene analysis. The majority of previous works on the topic focus on detecting monophonic and polyphonic acoustic events based on fully supervised data. However, it is difficult to make fully supervised acoustic event dataset since making strong labels is a very time-consuming process. In this paper, we propose an acoustic event detection framework for weakly supervised data which is labeled with only the existence of events. The model consists of a local detector and a global classifier. The local detector detects local audio words which contain distinct characteristics of events and the global classifier summarizes the information to make a decision on the recording. The experiments show that the proposed model has a lower EER on the CHiME Home dataset than other neural network based models.