ADReSS Dataset for Alzheimer’s Detection

Updated 1 July 2026

ADReSS dataset is a standardized benchmark corpus for Alzheimer’s detection through speech analysis, featuring demographically balanced clinical recordings.
The dataset includes high-quality manual and ASR transcriptions along with MMSE scores, supporting binary classification and regression evaluations.
Advanced feature extraction methods and fusion techniques from both acoustic and linguistic representations drive innovative machine learning applications in dementia research.

The Alzheimer’s Dementia Recognition through Spontaneous Speech (ADReSS) dataset is a standardized, demographically balanced benchmark corpus designed to enable rigorous comparative research on automatic Alzheimer’s Disease (AD) detection and related cognitive assessment tasks from speech and language. Created as part of the ADReSS Challenge, it offers high-quality, age- and sex-matched recordings and transcripts derived from a consistent clinical task: the “Cookie Theft” picture description from the Boston Diagnostic Aphasia Examination. ADReSS constitutes the de facto international standard for reproducible evaluation of machine learning, NLP, and speech processing methods for dementia detection, severity assessment, and model interpretability.

1. Dataset Structure, Composition, and Annotation

ADReSS comprises 156 English-speaking participants: 78 diagnosed with probable AD and 78 healthy controls (HC), with exact balance for age, gender, and, where available, basic cognitive status (e.g., MMSE 0–30). Each partition is demographically matched, ensuring no statistically significant bias by age or sex across AD and HC groups (Luz et al., 2020). The dataset is stratified into training (108 participants: 54 AD, 54 HC) and held-out test splits (48 participants: 24 AD, 24 HC); partitions are fixed and fully subject-independent. Typical participant ages range from mid-50s to late-80s, with mean MMSE for AD ≈17–20, for controls ≈29 (Luz et al., 2020, Sweidan et al., 24 Sep 2025).

All subjects perform a single, short (usually ≈30–60 s) “Cookie Theft” description; recordings are captured in controlled, quiet clinical environments at 16 kHz. Manual transcriptions follow the CHAT/CHILDES protocol, including explicit encoding of disfluencies, errors, retracing, unintelligible tokens (“xxx”), and pause annotations. Where specified, word-level timestamps are produced via forced alignment (Kaldi for manual transcripts; WhisperX for ASR) (Gómez-Zaragozá et al., 2023). The dataset further includes MMSE scores and standardized meta-data per participant (Sweidan et al., 24 Sep 2025).

Partition	Speakers	AD / HC	M / F	Audio Duration	MMSE Range
Train	108	54/54	48/60	~2 h	0–30
Test	48	24/24	22/26	~1 h	0–30
Total	156	78/78	70/86	~3 h	0–30

2. Prediction Tasks and Evaluation Protocols

ADReSS was constructed to support two principal tasks:

Binary AD Classification: Predict $\hat y \in \{0,1\}$ denoting HC or AD from each participant’s recording and/or transcript. Evaluation metrics are accuracy, precision, recall, and F1 score, with standard definitions (e.g., $F_1 = 2\frac{\pi\rho}{\pi+\rho}$ ) computed at the speaker level (Luz et al., 2020, Gómez-Zaragozá et al., 2023).
MMSE Regression: Predict the Mini-Mental State Examination (MMSE) total score (0–30) from spontaneous speech and/or transcript. The primary metric is root mean squared error ( $\mathrm{RMSE} = \sqrt{N^{-1}\sum_i(y_i-\hat y_i)^2}$ ), with test-set reporting on held-out subjects (Sarawgi et al., 2020).

Several works additionally report AUC (classification) or employ probability-calibrated outputs for threshold-free evaluation (Sweidan et al., 24 Sep 2025).

3. Feature Extraction: Speech, Language, and Fusion Approaches

ADReSS datasets allow both acoustic and linguistic representations. Acoustic features include the ComParE 2013 (6373-dim), emobase (988), eGeMAPS (88), and MRCG (6912) sets via openSMILE 2.1, with further functionals extracted per segment. Feature selection by Pearson correlation with utterance duration is standard (Luz et al., 2020).

Linguistic features exploit the richness of manual transcripts and include CLAN/EVAL-derived measures (e.g., mean length of utterance, type-token ratio, POS distributions), LIWC2015 (≈90 categories), and a variety of domain-driven syntactic, semantic, and psycholinguistic metrics (Balagopalan et al., 2020, Gómez-Zaragozá et al., 2023, Liu et al., 26 May 2025).

Recent work emphasizes powerful transfer and pre-trained representations: BERT-base and BERT-large for text (up to 1024-dim, typically last hidden layer; transcript treated as one sequence), and speaker embeddings (i-vectors, x-vectors) or self-supervised speech representations (wav2vec2, HuBERT, WavLM, Whisper) for audio (Pompili et al., 2020, Balagopalan et al., 2020, Gómez-Zaragozá et al., 2023, Liu et al., 26 May 2025, Liu et al., 4 Jun 2026). Automatic transcripts via ASR (18+ models including Whisper-large, Wav2Vec2, etc.) are now systematically evaluated, with WER as low as 26.4% for fine-tuned Whisper-large-v3 (Liu et al., 26 May 2025, Gómez-Zaragozá et al., 2023).

Fusion techniques include feature concatenation, attention-based multimodal pooling (e.g., cross-attention interpretability models (Liu et al., 26 May 2025)), and early/late classifier fusion (Pompili et al., 2020, Sarawgi et al., 2020). Sequence models (e.g., biLSTM with FastText embeddings) further exploit pause and punctuation mappings, showing that explicit pause encoding substantially boosts accuracy (Gómez-Zaragozá et al., 2023).

4. Modeling Pipelines and Performance Benchmarks

ADReSS research encompasses both classical and deep learning paradigms:

Classical ML: SVM with linear or RBF kernels, Random Forest, kNN, LDA, (decision trees, GPR, SVR for MMSE). Typical input is top-k selected hand-crafted features (Balagopalan et al., 2020, Luz et al., 2020, Gómez-Zaragozá et al., 2023).
Neural Pipelines: Bidirectional LSTM or attention-based models operating on FastText/BERT embeddings, with or without pause/punctuation augmentation (Gómez-Zaragozá et al., 2023, Balagopalan et al., 2020).
Prompted LLMs: Zero- and few-shot prompting on cleaned transcripts (e.g., Mistral-7B-Instruct), with “nested interleave” and MMSE-informed probability calibration (Sweidan et al., 24 Sep 2025).

Performance on the held-out 48-subject test set demonstrates consistent trends:

Pipeline	Input	Accuracy	F1	RMSE	Reference
Baseline SVM	Acoustic	0.625	0.62	6.14	(Luz et al., 2020)
SVM (linguistic)	Transcript	0.75	0.71–0.78	5.20	(Luz et al., 2020)
SVM + LIWC+pause	ASR	0.830	–	–	(Gómez-Zaragozá et al., 2023)
biLSTM+FT+pause	Manual	0.854	0.851	–	(Gómez-Zaragozá et al., 2023)
biLSTM+FT+pause	Whisper ASR	0.833	0.833	–	(Gómez-Zaragozá et al., 2023)
BERT finetune	Manual	0.833	0.83	4.56	(Balagopalan et al., 2020)
x-vectors+BERT	Fusion	0.8125	–	–	(Pompili et al., 2020)
LLM prompting	Manual	0.82	–	–	(Sweidan et al., 24 Sep 2025)
CoSTA/WavLM	Audio (aug)	0.8583	–	–	(Liu et al., 4 Jun 2026)
UA Ensemble	Multimodal	–	–	3.93	(Sarawgi et al., 2020)

Best results are generally achieved via deep linguistic models (BERT, biLSTM+FastText), carefully calibrated prompting with MMSE-informed probabilities, or audio-only augmentation using cognitively conditioned TTS synthesis (Liu et al., 4 Jun 2026, Gómez-Zaragozá et al., 2023, Balagopalan et al., 2020, Sweidan et al., 24 Sep 2025). Explicit pause encoding in transcripts and use of high-WER ASR transcripts can yield further gains, with some ASR-derived pipelines outperforming manual transcripts due to amplification of AD-specific lexical and structural markers (Liu et al., 26 May 2025, Liu et al., 4 Jun 2026).

5. Methodological Innovations and Interpretability

ADReSS fosters reproducible methodological advances across modalities:

Pause & Prosody Coding: Explicit textual encoding of pause length (e.g., “,” for <0.5 s, “.” for 0.5–2 s, “…” for >2 s) is highly predictive, elevating test accuracy by 4–8 points (Gómez-Zaragozá et al., 2023).
ASR Error Leveraging: Systematic evaluation of 36+ ASR transcript variants shows that certain ASR error profiles amplify informative differences, sometimes yielding higher classification accuracy than manual transcripts. These “useful blunders” preferentially shorten utterances or amplify repetition and truncation frequency in AD speech (Liu et al., 26 May 2025).
Attention-based Interpretability: Cross-attention models reveal which hand-crafted features align to pre-trained embedding representations; highest correlations are observed with features like repetition, content density, and syllable count (Liu et al., 26 May 2025).
Prompting Approaches: Few-shot prompting with MMSE-calibrated probabilities or reasoning-augmented prompts achieves near-SOTA accuracy and interpretable outputs, with per-exemplar estimates reflecting disease severity (Sweidan et al., 24 Sep 2025).

6. Design Principles, Limitations, and Future Directions

ADReSS’s design—age and sex matching, single elicitation task, consistent manual annotation—permits rigorous evaluation and method comparability, reducing demographic confounds and recording bias (Luz et al., 2020). Its limitations include modest sample size, focus on a single task (picture description), and the labor-intensive nature of manual CHAT transcription. The fixed subject-independent train/test split ensures experimental rigor but constrains model development in data-hungry paradigms.

Future research trajectories include: multimodal sequence and transformer fusion, longitudinal modeling of disease progression, robust ASR fine-tuning on pathological speech, and scaling to larger, more diverse cohorts (Pompili et al., 2020, Sarawgi et al., 2020, Liu et al., 4 Jun 2026). The impact of ADReSS extends beyond benchmarking, providing a foundation for clinical translation and for the reliable assessment of speech-based cognitive biomarker pipelines.

7. Impact and Benchmarking Significance

The ADReSS dataset has established itself as the reference standard for algorithmic comparison in Alzheimer’s speech technology research. It is the canonical testbed for classification pipelines based on manual and ASR-generated transcripts, interpretable and end-to-end audio models, MMSE regression architectures, and, more recently, LLM-based prompting and cognitively conditioned data augmentation (Luz et al., 2020, Gómez-Zaragozá et al., 2023, Liu et al., 4 Jun 2026, Sweidan et al., 24 Sep 2025). By anchoring empirical evaluation in a public, demographically controlled, and rigorously annotated dataset, ADReSS has accelerated methodological innovation and benchmarking accuracy for speech-based dementia detection research.