Paralinguistic singing attribute recognition using supervised machine learning for describing the classical tenor solo singing voice in vocal pedagogy

Humans can recognize someone's identity through their voice and describe the timbral phenomena of voices. Likewise, the singing voice also has timbral phenomena. In vocal pedagogy, vocal teachers listen and then describe the timbral phenomena of their student's singing voice. In this study, in order to enable machines to describe the singing voice from the vocal pedagogy point of view, we perform a task called paralinguistic singing attribute recognition. To achieve this goal, we first construct and publish an open source dataset named Singing Voice Quality and Technique Database (SVQTD) for supervised learning. All the audio clips in SVQTD are downloaded from YouTube and processed by music source separation and silence detection. For annotation, seven paralinguistic singing attributes commonly used in vocal pedagogy are adopted as the labels. Furthermore, to explore the different supervised machine learning algorithm for classifying each paralinguistic singing attribute, we adopt three main frameworks, namely openSMILE features with support vector machine (SF-SVM), end-to-end deep learning (E2EDL), and deep embedding with support vector machine (DE-SVM). Our methods are based on existing frameworks commonly employed in other paralinguistic speech attribute recognition tasks. In SF-SVM, we separately use the feature set of the INTERSPEECH 2009 Challenge and that of the INTERSPEECH 2016 Challenge as the SVM classifier's input. In E2EDL, the end-to-end framework separately utilizes the ResNet and transformer encoder as feature extractors. In particular, to handle two-dimensional spectrogram input for a transformer, we adopt a sliced multi-head self-attention (SMSA) mechanism. In the DE-SVM, we use the representation extracted from the E2EDL model as the input of the SVM classifier. Experimental results on SVQTD show no absolute winner between E2EDL and the DE-SVM, which means that the back-end SVM classifier with the representation learned by E2E as input does not necessarily improve the performance. However, the DE-SVM that utilizes the ResNet as the feature extractor achieves the best average UAR, with an average 16% improvement over that of the SF-SVM with INTERSPEECH's hand-crafted feature set.