Automatic Drum Transcription with Label Augmentation Using Convolutional Neural Networks.

Automatic drum transcription (ADT) aims to generate drum notes from drum music recordings. It is a significant subtask of music information retrieval. The successful transcription of drum instruments is a key step in the analysis of drum music. Existing systems use the target drum instruments as a separate training objective, which faces the problems of over-fitting and limited performance improvement. To solve the above limitations, this paper presents a label augmentation approach with the use of convolutional neural network via joint learning and self-distillation. Joint learning is introduced to gather drum and music information with the adoption of joint label and joint classifier. Self-distillation is proposed to use to enhance the expressive ability of the target model. We evaluate the designed system on well-known datasets and make a careful comparison with other ADT systems. The ablation studies are also conducted and a detailed analysis of the experimental results is provided. The experimental results show that the proposed system can achieve more competitive performance and outperform the state-of-the-art approach.