'CoVox': A matched vocal dataset for comparing singing and speech styles
The human voice is as diverse and individual as a fingerprint and can provide information about emotions, age, or health. In order to study vocal performances, researchers at the Max Planck Institute for Empirical Aesthetics (MPIEA) in Frankfurt am Main, Germany, have created a curated set of audio recordings with a total of 1,320 voice samples.
The dataset is freely available and has been validated in a published in Behavior Research Methods.
CoVox contains audio recordings of 22 Brazilian singers singing short melodies in three different styles—a lullaby, a pop song, and an opera aria. They also spoke the lyrics in two different styles: one as if they were addressing an adult and the other as if they were addressing a baby. This resulted in acoustic profiles that even laypeople could easily distinguish stylistically, the study found.
"What makes this dataset special is that it is fully matched: all the singing and speaking styles were performed by the same singers," explains first author Camila Bruder of the MPIEA.
"This consistency across performers makes it easier to compare the vocalizations across styles, making CoVox a controlled and directly comparable dataset."
"The dataset can be used as a source of experimental stimulus material for researchers to use in their own studies, but also as a subject of study in its own right, for example for comparisons between speech and singing," concludes Pauline Larrouy-Maestri, senior author at the MPIEA.
More information: Camila Bruder et al, CoVox: A dataset of contrasting vocalizations, Behavior Research Methods (2025).
Provided by Max Planck Society
