ADCeleb: A Longitudinal Speech Dataset from Public Figures for Early Detection of Alzheimer's Disease

ADCeleb encompasses 5347 recordings, amounting to approximately 25 hours of speech, from 80 celebrities, split evenly between 40 diagnosed with AD and 40 CNs. The speech samples were extracted from videos uploaded to YouTube. The dataset includes various accents (e.g., British, American, Canadian), professions, and age groups, annotated in the dataset. The AD and CN groups are matched based on age, sex, and ethnicity. This repository does not contain the actual audio recordings but provides metadata files with links to YouTube videos, speaker information, and transcriptions. You can utilize the following GitHub repository to download the audio files and segment recordings for each speaker based on their timestamps. Note: if you encounter problems or delays requesting access to the data, please send an email to [email protected] or any of the other authors). Key Features Diverse Acoustic Environments: the videos in the dataset feature a limited variety of complex multi-speaker acoustic environments, primarily including settings like calm studio interviews, public conferences, TV show talk delivered to large audiences, and informal talk in daily conversation. Publicly Available Data: all speech samples were extracted from publicly available videos uploaded to YouTube. Directory Structure Root directory The root directory (ADCeleb) contains two main subdirectories: AD: Speech data of celebrities diagnosed with Alzheimer’s disease. CN: Speech data of control subjects. Inside the root directory, you will also find the following key files: speakers_pairs.xlsx: contains the pairs of speakers (AD-CN) sex, age, and nationality-matched used in the classification. AD_demo.xlsx: is a metadata file with demographic information related to the AD group. CN_demo.xlsx: is a metadata file with demographic information related to the CN group. Speaker Folders Participants are organized into anonymized directories—each named either cn_xx (control subjects) or ad_xx (AD subjects), with xx running from 01 to 40. Inside every subject’s folder, you’ll find: metadata.csvA table of YouTube URLs pointing to that participant’s recordings, so you can download the raw audio or video files. Video-ID subfoldersEach recording is stored in its directory, named after the YouTube video identifier. Within each of those: Transcription File (.json): A word-level transcription, where every word is paired with its start and end timestamps. Speaker Timestamps File (.csv): it contains speaker labels and timestamps, indicating when each speaker is active during the recording. Column intersection: this column represents how much time the diarization segment and the transcript segment (or word) overlap. Column union: this column represents the total time covered by the diarization and transcript segments (or word). Target Speaker Annotation (speakers_info.csv): Column status: flags whether the folder’s subject is in the AD or control group. Column years_from_diagnosis: this column specifies how many years before or after diagnosis the recording took place. Column before_after_diagnosis: this column specifies whether the recording took place before or after diagnosis. For control subjects (CNs), we align their timelines to the diagnosis date of a matched AD participant, ensuring the two groups are comparable. Column video_context: this column specifies which type of video context the recording took place in (interviews, public conferences, informal talk and so on). Abstract Previous studies on early Alzheimer’s disease (AD) detection using speech have been limited by small sample sizes, scarce prodromal-phase recordings, and minimal longitudinal data. Many also rely on standardized tasks that do not reflect natural language use. To address these gaps, we introduce \textit{ADCeleb}, a longitudinal speech dataset comprising publicly available recordings from individuals who later disclosed an AD diagnosis. It includes samples from 40 individuals with prodromal AD and 40 cognitively normal controls (CNs), matched by age and sex, spanning 1 to 10 years before diagnosis. Classification experiments using multimodal models integrating speech and text achieved 0.72 accuracy in distinguishing AD from CNs in the 6- to 10-year pre-diagnosis window and 0.80 in the 1- to 5-year pre-diagnosis window. These results highlight the potential of speech-based technologies as non-invasive tools for early AD detection in real-world settings or for triage improvement in clinical trials. Reference Please acknowledge ADCeleb usage by using this reference in your publication: Gao, K., Favaro, A., Dehak, N., Velazquez, L.M. (2025) ADCeleb: A Longitudinal Speech Dataset from Public Figures for Early Detection of Alzheimer’s Disease. Proc. Interspeech 2025, 5688-5692, doi: 10.21437/Interspeech.2025-523. @inproceedings{gao25b_interspeech, title = {{ADCeleb: A Longitudinal Speech Dataset from Public Figures for Early Detection of Alzheimer’s Disease}}, author = {{Kunxiao Gao and Anna Favaro and Najim Dehak and Laureano Moro Velazquez}}, year = {{2025}}, booktitle = {{Interspeech 2025}}, pages = {{5688--5692}}, doi = {{10.21437/Interspeech.2025-523}}, issn = {{2958-1796}},}