Ensemble of Transformer and Convolutional Recurrent Neural Network for Improving Discrimination Accuracy in Automatic Chord Recognition

Automatic chord recognition is a task of recognizing and transcribing chords from music data such as popular music. Manual chord transcription requires highly technical knowledge and great effort. A chord is a typical musical feature. Realization of automatic chord recognition can enable their use for many purposes such as musical notation and structural analysis. For this reason, automatic chord recognition has become a major research task in the field of music information retrieval. In recent years, automatic chord recognition has widely used deep learning models. Convolutional Recurrent Neural Network (CRNN) and Transformer have achieved high accuracy. For this study, we focus on the differences in feature extraction approaches used by CRNN and Transformer, and propose an ensemble learning method using the two models. Additionally, we adopt an original overlap inference method to improve their accuracy by complementing the lack of temporal information. Results show that we achieved average accuracy of 78.92% under the traditional evaluation metrics, which are, respectively, 1.64% and 2.43% higher than those of CRNN and Transformer.