Reliability of automated and human transcriptions of non-Western music

Background: Cross-cultural analysis is essential to explore diversity and universality of music perception and cognition. Such analyses require symbolic representations of sounds such as score notation. However, transcription into notation is usually conducted by ear, which is time-consuming and subjective. Automated methods of music transcription might potentially solve these problems. Aims: The main objective is to evaluate the degree of agreement among human and automated transcriptions for a global song sample. To do so, we evaluated the degree of agreement among transcriptions created by human transcribers and the agreement among human transcribers and 10 automatic music transcription/vocal melody extraction algorithms. In particular, we created consensus transcriptions by merging the 3 individual transcriptions and compared the consensus transcriptions and the transcriptions contained in the Natural History of Song (Mehr et al., 2019) dataset for human-human agreement evaluation. Similarly, we compared the consensus transcriptions and machine transcriptions for the human-machine cases.

Jaar van publicatie:2021

Toegankelijkheid:Open