-
LoSST-AD is corpus of public interview transcripts that are recorded over several decades. There are 20 speakers in total, half of whom will receive Alzheimer’s disease diagnosis later in life. The aim of this corpus to allow assessing longitudinal Alzheimer’s disease related changes in spontaneous speech.
-
The name of this dataset is LoSST-AD, standing for Longitudinal Spontaneous Speech Transcripts for assessing changes in speech in Alzheimer’s Disease.
Composition
LoSST-AD includes anonymised transcripts of publicly available Youtube interviews of spontaneous speech, and the language features extracted from these transcripts.
There two datasets: original and capped. The original dataset consists of interviews from 20 individuals, half of whom receive Alzheimer’s disease diagnosis later in life. These speakers were paired with matched controls who have no known diagnosis of Alzheimer’s disease. Half of the speakers are male and half female. The original dataset consists of 135 transcripts and 222455 words.
The capped dataset is a sample of the original dataset consisting of 99 samples where the transcripts are shorter in length. The samples in the capped dataset include at least 3 transcripts from each speaker and the length in the transcripts is matched within a participant pair. The capped dataset was created to allow for the extraction of text-length-sensitive language features.
Both datasets include linguistic features extracted from these datasets.
The coding of the transcripts is as follows: The names of the transcripts from the individuals with AD start with 001-010, each number referring to a different speaker. The transcripts of the control group start with 011-020. The pairings of the AD and the control group speakers are 001-011, 002-012, 003-013 and so on. The next 3 digits in the file name refer to the year before diagnosis. For example, 001-03 refers to speaker 1 recorded 3 years before the diagnosis, and 002+02 refers to speaker 2 recorded 2 years after the diagnosis. As the control group speakers did not have a year of diagnosis, their last three digits are based on matching the age at recording time to their respective Alzheimer’s group speaker.
As the transcripts are anonymised, some information, such as proper names, family relations, and other identifiable information are changed or redacted. This may result in slight changes in the extracted language features. Some of the speech may have been unintelligible, resulting in errors or noise. The data may contain references to religious beliefs, speaker’s origin, occupation, and political opinion; however, effort has been put into anonymising the transcripts and making sure that this information is not linkable to the speaker’s identity. Anonymisation was further complicated by the public nature of the speakers, which often made it difficult to change all identity-related information.
The interviewer’s speech is not included in the transcripts, neither are direct quotes or song lyrics to allow assessing spontaneous speech of the speaker of interest.
Collection process
The data was acquired by transcribing publicly available Youtube interviews. The links to the interviews are not provided in an effort to protect the speakers’ identity. The data is in raw text format. The interviews were manually transcribed. Speaker diarisation has been applied, resulting in only the speech of the individuals of interest being transcribed and present in the dataset. Timestamps are not provided. The interviews were recorded between 1950 and 2021, and transcribed in 2020-2021.
The project was approved by the University of Cambridge Humanities and Social Sciences Research Ethics Committee.
The individuals were aware of the recordings taking place and being made public. As most individuals were deceased or public figures who are difficult to reach, it was not possible to notify them of the collection of this dataset or ask for consent. However, effort has been put into making the speakers unidentifiable.
Uses / Distribution / Maintenance
The data has been used to assess Alzheimer’s disease related changes in spontaneous speech. The following papers have used this dataset:
Petti, U., & Korhonen, A. (2024, May). LoSST-AD: A Longitudinal Corpus for Tracking Alzheimer’s Disease Related Changes in Spontaneous Speech. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024) (pp. 10813-10821).
Petti, U., Baker, S., Korhonen, A., & Robin, J. (2023). The Generalizability of Longitudinal Changes in Speech Before Alzheimer’s Disease Diagnosis. Journal of Alzheimer's Disease, (Preprint), 1-18.
Personally identifiable information has been changed or redacted following the guidelines for anonymisation. The dataset could be used for analysing longitudinal Alzheimer’s disease related changes in spontaneous speech. The dataset is available for research purposes on this website. The dataset will be improved and updated.
Ulla Petti is responsible for supporting, hosting, and maintaining this dataset (ump20 at cam.ac.uk)
The creation of the dataset is part of Ulla Petti’s PhD project, which was funded by the by the Economic and Social Research Council Cambridge Doctoral Training Partnership (DTP) grant number ES/P000738/1.
-
-
For any queries, please contact Ulla Petti
ump20 at cam.ac.uk