Multi-instrument recognition is the task of predicting the presence or absence of different instruments within an audio clip. A considerable challenge in applying deep learning to multi-instrument recognition is the scarcity of labeled data. OpenMIC is a recent dataset containing 20K polyphonic audio clips. The dataset is weakly labeled, in that only the presence or absence of instruments is known for each clip, while the onset and offset times are unknown. The dataset is also partially labeled, in that only a subset of instruments are labeled for each clip. In this work, we investigate the use of attention-based recurrent neural networks to address the weakly-labeled problem. We also use different data augmentation methods to mitigate the partially-labeled problem. Our experiments show that our approach achieves state-of-the-art results on the OpenMIC multi-instrument recognition task.
View on arXiv