SPED: A Sight-singing Dataset for Performance Evaluation

Abstract

Sight-singing is the process when a learner reads and sings a musical score at the same time. Most of the existing music datasets focus on note transcription tasks, only containing audio and pitch annotation. In this paper, we propose the first large-scale sight-singing dataset for performance evaluation named SPED, which contains 1,011 musical scores and 304,368 sight-singing recordings sung by a total of 9,000 professional, semiprofessional and inexperienced learners. The dataset focuses on learners’ sight-singing performance data, and records the pitch and rhythm score sequences of the recordings, as well as the key signature, beat, and other music score information. The dataset can be used to analyze learners’ sight-singing ability, score difficulty, and can be applied in knowledge tracing tasks. The dataset is statistically analyzed and results reflect a good consistency with general knowledge in the field of sight-singing, which indicates that the data quality meets expectations.