On-Line Chord Recognition Using FifthNet with Synchrosqueezing Transform
Rikuto Ito, Natsuki Akaishi, Kohei Yatabe, Yasuhiro Oikawa · 2023
Automatic chord recognition is a fundamental and important task in music information processing. FifthNet is a recently proposed chord recognition framework based on a deep neural network (DNN). Its aim is to reduce the computational and memory loads by taking advantage of knowledge on music signals. Since FifthNet achieved the state-of-the-art performance and requires less computational resource compared to the other DNN-based methods, it seems suitable for real-time applications. However, the original FifthNet cannot be directly used in real-time because it applies the spectral reassignment method in its feature extractor. The reassignment method requires some look-ahead frames that results in unavoidable latency. In this paper, we propose to replace the reassignment method of FifthNet with the synchrosqueezing transform to reduce the amount of required look-ahead frames and the computational requirement. Our modification makes FifthNet on-line and removes the obstacle towards real-time execution. In addition, the experimental result shows that the accuracy of chord recognition can be improved by the proposed modification.