Korean singing dataset includes detailed vocal expression labels
Summary and headline written by AI from the source article. How we work
Researchers created Gugak-VocalSet, a new dataset of 24.5 hours of Korean traditional singing recordings from the National Gugak Center in South Korea.
Twenty professional singers performed 305 solo tracks in a studio, maintaining a consistent key and tempo. The dataset includes not only audio and lyrics, but also 68,000 labeled segments detailing sigimsae, a system of pitch contours that express emotion in Korean vocal music.
Fourteen experts in Korean traditional music annotated the recordings using a standardized system designed to ensure consistent labeling across different musical genres. Experiments using the dataset showed that representations focused on fundamental frequency performed well in classifying sigimsae, and that both pitch and spectral cues help to identify specific vocal events. The creators made the dataset publicly available, hoping to support further research into Korean traditional music and vocal expression.
