Classification of Vocal Type in Choir Using Convolutional Recurrent Neural Network (CRNN)

Determining vocal types is a crucial aspect that forms the foundation of choir composition. It requires an in-depth understanding of music. This fact makes the construction of a vocal type classification model in a choir a multidisciplinary experiment. In addition to the technical review, a profound musicality review is also needed. Consequently, determining vocal types in a choir cannot be done merely based on rules regarding the range of an individual's voice. This research focuses on classifying vocal types using features other than the fundamental frequency (f0), such as timbre and vocal stability. In this study, data were collected from the members of the ITB Student Choir (PSM-ITB), with the labeling process carried out by the PSM-ITB vocal training team. The last research, that uses Convolutional Neural Network, gave an accuracy of 0.74. This shows that the research still has room for improvement. CNN can use the feature of timbre to predict the class but didn't consider the stability of the audio. On the contrary, Recurrent Neural Network gave an accuracy of 0.65 which is less than satisfactory. RNN can use the feature of stability, but the timbre feature is not one that taken into account. This study provides relatively good results compared to the previous models. The Convolutional Recurrent Neural Network model achieved an accuracy of 0.87, higher than the existing Convolutional Neural Network model and the Recurrent Neural Network model.