Weakly Labeled Learning Using BLSTM-CTC for Sound Event Detection

Abstract

In this paper, we propose a method of weakly labeled learning of bidirectional long short-term memory (BLSTM) using connectionist temporal classification (BLSTM-CTC) to reduce the hand-labeling cost of learning samples. BLSTM-CTC enables us to update the parameters of BLSTM by loss calculation using CTC, instead of the exact error calculation that cannot be conducted when using weakly labeled samples, which have only the event class of each individual sound event. In the proposed method, we first conduct strongly labeled learning of BLSTM using a small amount of strongly labeled samples, which have the timestamps of the beginning and end of each individual sound event and its event class, as initial learning. We then conduct weakly labeled learning based on BLSTM-CTC using a large amount of weakly labeled samples as additional learning. To evaluate the performance of the proposed method, we conducted a sound event detection experiment using the dataset provided by Detection and Classification of Acoustic Scenes and Events (DCASE) 2016 Task 2. As a result, the proposed method improved the segment-based F1 score by 1.9% compared with the initial learning mentioned above. Furthermore, it succeeded in reducing the labeling cost by 95%, although the F1 score was degraded by 1.3%, comparing with additional learning using a large amount of strongly labeled samples. This result confirms that our weakly labeled learning is effective for learning BLSTM with a low hand-labeling cost.