Celebs4PD
Celebs4PD at a glance
Overview
Dataset composition
| Group | Participants | Recordings | Duration (h) | Size (GB) |
|---|
Spoken language
Share of the 700 recordings in each language. Two participants have recordings in both languages.
Recording coverage per participant
SD is the sample standard deviation. Distributions are right-skewed, so medians are also shown.
Recording coverage per participant pair
Hours and recordings per pair
Hours of target-speaker audio per participant, one row per PD–Control pair (pairs 1–15 male, 16–30 female). PD participants contribute 64.96 h and Control participants 73.68 h. Hover a bar for the pair total.
Audio per pair ranges from 1.4 h (pair 11) to 10.4 h (pair 24), which reflects data availability at the time of dataset curation. Only 10 of 30 pairs fall within 3–5 h, around the 4.6 h mean.
Participant characteristics
Age group at time of recording
Each bar shows the number of recordings made while the participant was within that age group.
Age at reported year of diagnosis
Each point is one participant, and the curve is a kernel density estimate of these ages.
Profession by diagnostic group
13 categories, sorted by total. Overlapping points indicate equal counts. In points with both sexes, the solid wedge is the female share.
Race by participant pair
One square per participant, by sex and diagnostic group. Mirrored squares are PD–Control partners, who share race in 28 of 30 pairs.
Nationality by participant pair
One square per participant, by sex and diagnostic group. Mirrored squares are PD–Control partners, who share nationality in 16 of 30 pairs.
Recording timeline by participant pair
Each point is one recording, placed by its year relative to baseline: −3 is three years before, +3 three years after. Baseline is the reported year of diagnosis for each PD participant and, for each Control, the year of diagnosis of their matched PD partner (Control row n with PD row n). Pairs 1–15 are male and 16–30 are female.
Key features
How to split this dataset
This dataset is designed to be split by participant pairs, which preserves demographic balance, limits the influence of potential confounders, and prevents speaker leakage. The split explorer helps you visualize and compare training, validation, and test configurations, and choose the participant pairs best suited to your research.