PaperlayerView exact revision on arXiv

Source-linked AI summary

Acoustic Biomarkers for Alzheimer's Disease Detection from Spontaneous Speech

Anh L Tran, Arjee Javellana Restar, Bereket Kindo

arXiv:2610.12500v1eess.AScs.LGcs.SD

TL;DR

The paper asks whether acoustic features from spontaneous speech can provide useful signals for AD-related clinical decision support. It evaluates machine-learning models on Cookie Theft speech recordings and reports that acoustic descriptors carry meaningful information, with an ALBERT–Whisper stack achieving the lowest compared log loss.

  • Problem

    The study examines whether audio-derived measurements can contribute useful, explainable signals for AD clinical decision support without replacing clinical diagnosis.

    Source: §1 · p. 1

  • Method

    The authors extract acoustic descriptors from standardized picture-description recordings and compare machine-learning models across acoustic, sequential, transcription, and raw-audio representations.

    Source: §1 · p. 2, §3.3 · p. 3

  • Results

    0.6580 log loss for ALBERT stacked with Whisper was the lowest in the model comparison; the study also reports meaningful information in spontaneous-speech acoustic features.

    Source: §4 · p. 4, §6 · p. 9

  • Takeaways & Limitations

    The findings support further development of non-invasive, scalable speech-based screening tools.

    Source: §6 · p. 9

    The dataset is moderate in size, acoustic markers vary with environment, recording conditions, and demographics, and speech impairment is not specific to AD.

    Source: §5 · p. 9

Abstract

from arXiv · show

Alzheimer's disease (AD) remains a major public health challenge, and early identification is crucial for timely intervention and care planning. Recent work has shown that speech-based digital biomarkers can capture subtle changes in vocal production, prosody, and timing that often precede or accompany cognitive decline. In this study, we investigate whether acoustic features extracted from spontaneous speech can reliably distinguish individuals with AD from healthy controls. Using a structured speech dataset derived from the Cookie Theft picture-description task, we extract clinically informative low-level descriptors spanning prosodic, spectral, and temporal domains, and evaluate a range of machine learning classifiers. Our results indicate that acoustic biomarkers carry strong discriminative value, with the best-performing models achieving clinically meaningful separation between diagnostic groups. These findings support the use of non-invasive, low-cost audio-based screening tools and motivate further work on multimodal and longitudinal assessment frameworks for dementia risk detection.

1 Introduction

The paper examines whether acoustic measurements from spontaneous speech can contribute useful, explainable signals for AD clinical decision support. It frames speech as an accessible source of indicators of cognitive change and evaluates machine-learning classification using a standardized picture-description task.

  • Speech is non-invasive and inexpensive to capture, while traditional diagnostic pipelines can be costly, time-consuming, and inaccessible for routine screening.The introduction also notes that speech is sensitive to changes in cognition, executive function, and motor control.Source: §1 · p. 1
  • AD-related changes in lexical access, fluency, rhythm, and prosody can be measured through acoustic features including pitch variability, formants, pauses, and spectral structure.Source: §1 · p. 1
  • The study tests acoustic-only measurements from a standardized picture-description task as explainable signals for clinical decision support, not as a replacement for diagnosis.Machine-learning models are trained to distinguish patients with AD from controls.Source: §1 · pp. 1–2

2 Related work

Prior work links cognitive impairment and dementia to changes in speech production, including acoustic and temporal patterns that can be analyzed without transcription. This study evaluates an interpretable, signal-derived modeling pipeline for spontaneous speech.

  • Prior studies associate dementia with reduced fluency, more hesitation and longer pauses, and altered prosody; acoustic analysis can distinguish AD from healthy aging without linguistic transcription.Source: §2 · p. 2
  • Dementia-relevant acoustic descriptors span frequency, amplitude, and spectral properties, motivating compact features extractable from short audio samples.Examples include pitch and formant trajectories, loudness and energy, MFCCs, spectral slope, and HNR.Source: §2 · p. 2
  • Studies of Cookie Theft datasets report predictive value in purely acoustic features, alongside repeatedly associated articulation-rate, pause-duration, vocal-stability, and cepstral measures.Source: §2 · p. 2
  • The study evaluates interpretable, signal-derived descriptors and classification models for spontaneous speech, avoiding reliance on manual transcription and extensive annotation.The stated practical motivation is scalable screening without expensive annotation or expert transcription workflows.Source: §2 · p. 2

3 Materials and Methods

The methods use Cookie Theft picture-description recordings labeled by cognitive group, extract low-level acoustic features, and compare models across several speech representations.

3.1 Dataset

  • The curated Boston Cookie Theft dataset contains spontaneous recordings labeled as normal cognition, MCI, or ADRD for evaluating discrimination among the groups.Source: §3.1 · pp. 2–3

3.2 Preprocessing and feature extraction

  • Recordings were standardized and inspected before extracting low-level prosodic, energy, spectral, temporal, and frame-dynamics descriptors.The feature set includes pitch and contour measures, MFCCs and formant-related features, pause and voiced/unvoiced durations, and speech-rate proxies.Source: §3.2 · p. 3

3.3 Modeling framework

  • Models were evaluated on aggregated and sequential acoustic features, transcriptions, raw-audio embeddings, and stacked combinations, using multiclass log loss as the primary metric.The study also computed SHAP values for the best aggregated-acoustic model.Source: §3.3 · p. 3

4 Experiments and Results

The lowest reported log loss came from stacking ALBERT with Whisper, while SHAP analyses identified influential acoustic descriptors and differing feature patterns across diagnostic groups.

  • 0.6580 log loss for ALBERT stacked with Whisper was the lowest, versus 0.7404 for XGBoost stacked with TF-IDF and 0.7430 for Whisper alone.Lower log loss indicates predicted probabilities closer to the true labels.Source: §4 · pp. 3–4, Figure 1 · p. 4
  • 0.8299 log loss made XGBoost the strongest aggregated-acoustic classifier, ahead of CatBoost at 0.8530 and random forest at 0.8989.Logistic regression scored 0.9900, and the decision tree scored 2.7220, described as an outlier consistent with overfitting.Source: §4 · p. 4
  • 0.7430 log loss made Whisper the best single model; ALBERT scored 0.8270, compared with TF-IDF at 0.8634 and Word2Vec at 0.9251.The bidirectional LSTM scored 0.8646 versus 0.8879 for the unidirectional LSTM.Source: §4 · p. 4
  • The top pooled SHAP contributors included equivalent-sound-level variability, age, F0 slopes, loudness peaks, spectral features, and pause and voicing measures.The reported pattern gives greater discriminative weight to loudness variability, pitch dynamics, spectral shape, and timing than to any single speech-rate summary.Source: §4 · pp. 4–5, Figure 2 · p. 5
  • Class-specific rankings differed: pitch-slope and loudness-peak features dominated for normal cognition, while sound-level variability, F0 slope, and unvoiced-segment length stood out for MCI.Source: §4 · pp. 5–7, Figure 3 · p. 6, Figure 4 · p. 7
  • For ADRD, age ranked highest, with unvoiced alpha ratio, spectral flux, loudness, and voicing statistics also contributing.Source: §4 · pp. 5, 8, Figure 5 · p. 8

5 Discussion

The discussion identifies physical and temporal vocal changes as candidate markers of cognitive impairment, highlights interpretable low-dimensional features, and cautions that dataset and speech-specificity constraints limit clinical use.

5 Discussion

  • Pitch instability, altered energy patterns, more frequent pauses, and greater vocal-tract-dynamics variability are candidate markers of cognitive impairment.Source: §5 · p. 9
  • Interpretable, low-dimensional descriptors retain predictive value without treating audio as a black box, supporting their relevance to clinical translation.Source: §5 · p. 9
  • Moderate dataset size, recording and demographic influences, and speech changes shared with other conditions mean the model is decision support, not stand-alone diagnosis.Source: §5 · p. 9
  • The proposed direction is multimodal fusion of acoustic features with linguistic, cognitive, and patient metadata to improve robustness and clinical utility.Source: §5 · p. 9
  • Future work should use larger, more diverse datasets, cross-site external validation, and longitudinal analyses to assess progression tracking and early detection.Source: §5 · p. 9

6 Conclusion

The conclusion reports that spontaneous-speech acoustic features provide meaningful information for Alzheimer's disease detection and may contribute to digital clinical assessment. It points to screening and monitoring opportunities while calling for broader validation and longitudinal work.

6 Conclusion

  • Acoustic features extracted from spontaneous speech provide meaningful information for Alzheimer's disease detection.Source: §6 · p. 9
  • Standard machine-learning methods use these features to capture clinically relevant differences in voice production, prosody, and temporal structure.Source: §6 · p. 9
  • Speech biomarkers may contribute to digital clinical assessment and low-cost screening and monitoring, with future work focused on multimodal integration, cross-population validation, and longitudinal tracking.Source: §6 · p. 9
Loading 2610.12500v1…