Note-level Automatic Guitar Transcription Using Attention Mechanism
Sehun Kim, Tomoki Hayashi, Tomoki Toda · 2022 30th European Signal Processing Conference (EUSIPCO) · 2022
We propose a method that effectively generates a note-level transcription from a guitar sound signal. In recent years, there have been many successful guitar transcription systems. However, most of them generate a frame-level transcription rather than a note-level transcription. Furthermore, it is usually difficult to effectively model long-term characteristics. To address these problems, we propose a novel model architecture using an attention mechanism along with a convolutional neural network (CNN). Our model is capable of modeling both short-term and long-term characteristics of a guitar sound signal and a corresponding guitar transcription. A beat-informed quantization is implemented to generate a note-level transcription. Furthermore, multi-task learning with frame-level and note-level estimations is also implemented to achieve robust training. We conducted experimental evaluations on our method using a publicly available acoustic guitar dataset. We confirmed that 1) the proposed method significantly outperforms the conventional method based on a CNN in frame-level estimation performance and that 2) the proposed method can also generate note-level guitar transcription while preserving high estimation performance.