Naturalistic Daily Voice Sampling at Home for CNS Digital Biomarker Research

This paper proposes a research framework. It does not claim that daily home voice sampling is clinically superior to standardized clinic assessment, nor that it currently constitutes a validated diagnostic, prognostic, or regulatory endpoint.

The central claim is narrower: daily naturalistic voice conversations at home may provide complementary, ecologically valid, longitudinal data for CNS digital biomarker research, and this hypothesis should be tested through fit-for-purpose validation studies.

Speech and voice are increasingly studied as candidate digital biomarkers for central nervous system disorders, including Alzheimer's disease, mild cognitive impairment, Parkinson's disease, depression, and other neuropsychiatric conditions. Much of the field still relies on structured speech tasks administered in clinical or research settings. Such data are valuable because they are standardized, anchored to clinical workflows, and easier to compare across participants.

Clinic-based speech samples may also be affected by assessment context: participant anxiety, novelty effects, unusually high effort, fatigue from travel, interaction with unfamiliar staff, time-of-day effects, and the artificiality of performing under observation.

This white paper proposes a testable hypothesis: daily or high-frequency voice data collected through natural conversations in the participant's home environment may produce complementary data that are more ecologically valid and longitudinally sensitive than isolated clinic-based speech samples alone. The hypothesis is grounded in ecological momentary assessment, remote digital cognitive assessment, decentralized clinical trial methodology, and emerging speech biomarker research.^[1][2][3]

The proposed model does not replace controlled assessments. Instead, it integrates them: standardized clinic-based speech tasks provide clinical anchors, while repeated home conversations provide longitudinal, real-world behavioral signal. If validated, this hybrid model could support CNS trials by improving intra-individual baselining, state-trait separation, detection of subtle functional change, and measurement of real-world communication.

The central premise of this paper is that CNS-relevant speech and language changes are often dynamic, context-sensitive, and longitudinal. A single clinic-based speech sample may be clinically useful, but it is unlikely to fully characterize a participant's everyday communicative function, affective state, cognitive effort, or intra-individual variability.

We propose that repeated voice sampling through natural telephone conversations at home can capture a different class of signal: not merely task performance, but real-world verbal behavior over time. This distinction matters for digital biomarker research. A biomarker derived from a one-time standardized task may estimate ability under controlled conditions. A biomarker derived from daily naturalistic conversations may estimate functional communication under lived conditions.

Daily or high-frequency naturalistic voice recordings collected at home may provide more ecologically valid and longitudinally sensitive data for CNS digital biomarker research than isolated clinic-based speech assessments alone, while still requiring clinic-based anchors for validation and interpretation.

This hypothesis has four components:

• Ecological validity: home conversations may better reflect everyday speech, cognition, affect, and functional communication.
• Longitudinal sensitivity: repeated sampling may estimate personal baseline, variability, and trajectory more accurately than one or two clinic visits.
• Assessment-artifact reduction: familiar repeated conversations may reduce stress, novelty effects, and performance distortion relative to formal assessment settings.
• Signal enrichment: natural dialogue may reveal features that structured tasks miss, such as topic maintenance, conversational repair, spontaneous narrative coherence, affective tone, and social reciprocity.

CNS clinical research is constrained by noisy outcomes, heterogeneous disease trajectories, variable symptom expression, and limited sensitivity of episodic assessments. Cognitive and neuropsychiatric symptoms fluctuate across days and contexts. Mood, sleep, medication timing, stress, fatigue, pain, social interaction, and time of day can influence performance. Yet many trials still rely heavily on periodic assessments that compress this variability into a small number of visits.

This problem is not unique to speech. It is a general challenge in CNS measurement. However, speech is especially sensitive to context because it reflects multiple systems at once: language, memory, executive function, processing speed, affect, motor planning, respiratory coordination, hearing, social motivation, and interpersonal comfort. A participant's speech during a formal research visit may differ from their speech during an ordinary conversation at home.

The practical consequence is that clinic-based speech assessments may have strong internal standardization but weaker ecological validity. Conversely, home-based naturalistic conversations may have weaker experimental control but stronger real-world relevance. The research opportunity is to combine both rather than treat them as substitutes.

3.1 Ecological momentary assessment and natural environments

Ecological momentary assessment provides a direct methodological precedent. Shiffman, Stone, and Hufford define EMA as repeated sampling of current behaviors and experiences in real time and in natural environments. EMA was developed to reduce recall bias, improve ecological validity, and study processes as they unfold in everyday life.^[1]

Daily home voice sampling can be understood as an EMA-adjacent approach applied to speech and language. Instead of asking participants to retrospectively report how they felt or functioned, the system observes repeated speech behavior in context. The resulting data may allow researchers to model micro-variation, weekly patterns, adaptation effects, and deviations from personal baseline.

3.2 Remote digital health technologies in clinical investigations

Regulatory and clinical trial practice is increasingly open to remote data acquisition. FDA guidance on Digital Health Technologies for Remote Data Acquisition in Clinical Investigations provides recommendations for using hardware and software systems to acquire data remotely from participants, noting potential improvements in trial efficiency and participant convenience.^[2]

This does not imply that any remote measure is valid. It means that remote acquisition is compatible with clinical investigation when the technology, measure, and context of use are appropriately specified and validated. For naturalistic voice biomarkers, the relevant question is not simply whether audio can be collected remotely, but whether the derived measure is accurate, reliable, interpretable, clinically meaningful, and fit for the intended use.

3.3 Remote and unsupervised digital cognitive assessment

A 2025 scoping review in npj Digital Medicine concluded that remote and unsupervised digital assessments can improve scalability, measurement reliability, and ecological validity in preclinical Alzheimer's disease research, enabling the capture of subtle cognitive changes. This supports the broader principle that repeated remote measurement may be valuable when subtle change, ecological validity, and scalability are critical.^[4]

The same logic applies to speech-based measures. Many CNS changes may first appear as small shifts in processing speed, fluency, word retrieval, coherence, affective tone, or conversational initiative. These changes may be difficult to detect from one clinic visit but more visible across repeated samples.

3.4 Speech as a CNS-relevant signal

Speech is a plausible CNS digital biomarker source because it is generated by distributed cognitive, affective, linguistic, and motor systems. Systematic review evidence suggests that speech-based biomarkers have diagnostic utility for distinguishing mild cognitive impairment from cognitively unimpaired status, while also emphasizing heterogeneity and the need for stronger validation.^[5]

Automated speech analysis studies commonly extract acoustic features such as speech rate, pause duration, articulation rate, pitch variability, intensity, and voice quality, as well as linguistic features such as lexical diversity, syntactic complexity, semantic coherence, repetition, pronoun use, information density, and narrative structure. Naturalistic conversations may expand this feature space by adding turn-taking, topic maintenance, conversational repair, spontaneity, affective valence, and social reciprocity.

3.5 Momentary affect and state-sensitive speech

Speech is not only a trait marker. It also varies with momentary affective state. A 2024 ambulatory assessment study in JMIR Mental Health found that pitch variability, speech pauses, and speech rate were associated with depression severity, positive affect, valence, and energetic arousal. This supports the idea that high-frequency speech sampling may be useful not only for disease classification but also for capturing state variation.^[6]

For CNS trials, this is methodologically important. A single speech sample may be confounded by temporary state. Daily sampling can model state variation rather than treating it as random noise.

3.6 Telephone-based speech collection in older adults

Telephone-based speech collection is especially relevant for older adults because it avoids requiring an app, wearable, camera, screen interaction, or advanced digital literacy. The SPeAk study reported that a brief telephone-based cognitive and speech assessment was feasible and acceptable among people at risk for Alzheimer's disease dementia. A separate study found that collecting speech samples by telephone was well tolerated, practical, inexpensive, and capable of producing good-quality data for natural language processing.^[7][8]

This matters for implementation. A method that is theoretically valid but operationally burdensome may fail in real-world elderly populations. The telephone may be a strong interface precisely because it is familiar, low-friction, and already embedded in daily life.

3.7 Validation standards for digital measures

The Digital Medicine Society's V3 framework provides a useful structure for evaluating digital measures: verification of the sensor or data capture component, analytical validation of the algorithm, and clinical validation of the outcome measure for the intended context of use. DiMe has also described V3+ as extending the framework with usability validation for digital clinical measures.^[3][9]

For daily naturalistic voice sampling, this framework is essential. The data acquisition pipeline must be technically verified, algorithms must be analytically validated against reference labels or benchmark tasks, and derived measures must be clinically validated against meaningful CNS outcomes before being used as endpoints or decision-support tools.

The proposed model is a hybrid measurement architecture: standardized clinic assessments provide anchors; daily home conversations provide longitudinal ecological signal.

This architecture intentionally avoids a false binary. Clinic data and home data answer different questions. Clinic data asks: how does the participant perform under standardized conditions? Home data asks: how does the participant communicate and vary in everyday life?

5.1 Primary research question

Do daily or high-frequency naturalistic voice conversations collected at home provide ecologically valid and longitudinally sensitive speech-derived data for CNS digital biomarker research, beyond what is captured by isolated clinic-based speech assessments?

5.2 Secondary research questions

• Do home-based voice features correlate with validated measures of cognition, mood, function, or disease severity?
• Do longitudinal voice features predict future clinical change better than baseline clinic speech features alone?
• Does repeated familiarity with an AI caller reduce novelty effects and assessment-related stress?
• Which features are stable enough for trait-level measurement, and which are state-sensitive?
• Can within-person change improve sensitivity compared with cross-sectional population norms?
• Can naturalistic voice data support exploratory or secondary endpoints in CNS trials?

5.3 Hypotheses

A lab-ready program should be staged. The goal is not to jump directly to clinical efficacy claims, but to progressively establish feasibility, reliability, validity, and fit-for-purpose utility.

6.1 Phase 0 - Technical feasibility and protocol optimization

Objective: determine whether daily telephone voice collection is technically feasible in the intended population and whether the audio/transcript quality is sufficient for feature extraction. Sample: 20-40 older adults across cognitive status strata. Duration: 2-4 weeks.

Measures should include call completion, audio quality, transcription quality, participant burden, technical failures, time-of-day effects, and consent comprehension. Outputs include finalized protocol, data dictionary, QC thresholds, adverse event procedure, and participant-facing consent materials.

6.2 Phase 1 - Observational longitudinal cohort

Objective: characterize feature stability, within-person variability, adherence, acceptability, and associations with clinical anchors. A reasonable sample would include 100-200 participants across cognitively healthy controls, subjective cognitive decline, MCI, early dementia, depression, Parkinson’s disease, or another CNS cohort depending on target indication.

Duration should be 8-12 weeks minimum, and ideally 6 months for progression signal. Data should include daily or near-daily naturalistic phone conversations, baseline and follow-up clinic assessments, structured speech tasks, mood scales, cognitive scales, functional measures, and optional caregiver reports.

6.3 Phase 2 - Within-subject clinic vs home comparison

Objective: isolate the effect of environment, task type, and sampling frequency. Each participant provides clinic structured speech, home structured speech, home naturalistic speech, and repeated home samples over time. Analysis compares reliability, ecological validity, clinical correlation, participant stress, and predictive value across conditions.

6.4 Phase 3 - Embedded exploratory endpoint in a CNS trial

Objective: evaluate whether daily naturalistic voice features can function as exploratory or secondary digital endpoints in a clinical trial. Potential use cases include Alzheimer’s disease, MCI, Parkinson’s disease, late-life depression, major depressive disorder, multiple sclerosis, post-stroke cognitive impairment, or neurodegenerative disease trials.

The primary test is whether adding daily voice trajectories improves prediction of clinical change, responder status, functional decline, or adverse neuropsychiatric events. The regulatory posture should remain exploratory until sufficient verification, analytical validation, clinical validation, and usability validation are established.

7.1 Endpoint hierarchy

7.2 Speech and language feature domains

7.3 Trait, state, and change features

A serious analysis should separate three classes of features:

• Trait-like features: relatively stable characteristics such as baseline speech tempo, typical lexical richness, or long-term prosodic profile.
• State-sensitive features: temporary changes associated with mood, fatigue, sleep, stress, illness, or medication timing.
• Change-over-time features: progressive deviations from personal baseline, such as increasing pauses, declining coherence, reduced narrative detail, or reduced conversational initiative.

A lab-ready study should prespecify the analysis plan before data lock. The following plan is suitable for a feasibility-to-validation program.

8.1 Data preprocessing and quality control

• Define minimum audio duration, signal-to-noise thresholds, transcription confidence thresholds, and exclusion criteria.
• Segment calls into comparable units: entire call, first three minutes, prompted response windows, or topic-specific segments.
• Track device, phone line quality, time of day, background noise, participant-reported sleep/fatigue, and medication timing where feasible.
• Separate automated features from human-coded validation subsets.

8.2 Reliability analysis

• Estimate within-person and between-person variance components using mixed-effects models.
• Use intraclass correlation coefficients to estimate reliability of each feature.
• Estimate the number of daily samples needed to reach acceptable reliability for each feature class.
• Assess test-retest reliability for structured tasks and naturalistic segments separately.

8.3 Clinical association analysis

• Test associations between voice features and clinical anchors using regression models adjusted for age, sex, education, language, hearing status, baseline cognitive status, depression, and device quality.
• Use mixed-effects models to test whether within-person change in voice features predicts within-person change in clinical measures.
• Compare clinic-only models against clinic-plus-home models using out-of-sample prediction, likelihood ratio tests, and calibration metrics.
• Avoid overclaiming cross-sectional classification accuracy unless externally validated.

8.4 Predictive modeling

• Use nested cross-validation and participant-level train/test separation to prevent leakage.
• Prespecify primary model families before analysis: mixed-effects regression, penalized regression, gradient boosting, or time-series models.
• Use interpretable features where possible before high-dimensional black-box modeling.
• Validate against future clinical change, not only current diagnostic category.

8.5 Missing data and adherence

• Model missingness explicitly; missed calls may be informative rather than random.
• Distinguish technical missingness from participant refusal, fatigue, hospitalization, distress, or cognitive decline.
• Use sensitivity analyses to test whether results are robust to missing-not-at-random assumptions.
• Report adherence by cohort, cognitive status, age, language, and technology comfort.

8.6 Multiplicity and exploratory features

Speech analysis can generate hundreds or thousands of candidate features. A credible protocol should define a limited primary feature set and treat high-dimensional discovery as exploratory. False discovery control, dimensionality reduction, preregistered feature families, and external replication are essential to avoid spurious biomarker claims.

Daily home speech data should be validated through a V3/V3+ style framework.

The most common failure mode in digital biomarker programs is collecting large volumes of data before proving that the derived measures are reliable, clinically interpretable, and fit for purpose. This program should reverse that order: define the context of use first, then validate the measure against that use.

Voice data collected in daily conversations is highly sensitive. It may contain health information, family relationships, financial concerns, emotional disclosures, names, addresses, and incidental third-party information. For older adults and cognitively vulnerable individuals, the ethical threshold is high.

• Consent: participants must understand that they are speaking to an AI system and that voice data may be recorded and analyzed for research.
• Capacity: protocols should include procedures for cognitive vulnerability, surrogate consent where appropriate, and ongoing assent.
• Privacy: recordings, transcripts, features, and summaries should be minimized, encrypted, access-controlled, and governed by a clear retention policy.
• Non-diagnostic positioning: until validated, the system should not diagnose depression, dementia, Parkinson’s disease, or any other CNS disorder.
• Human oversight: concerning signals should trigger human review, not automated clinical conclusions.
• Bias: performance should be evaluated across accents, dialects, languages, education levels, hearing status, cognitive status, sex, and cultural communication styles.
• Psychological safety: repeated conversations may surface distress, grief, confusion, or sensitive memories; protocols should define escalation and discontinuation procedures.

A credible research paper must state how the hypothesis could be wrong.

• Home audio may be too noisy or variable to support reliable feature extraction.
• Natural conversation may be too topic-dependent to permit valid longitudinal comparisons.
• Participant adherence may decline over time, especially in cognitively impaired groups.
• The AI interaction itself may alter speech behavior, creating intervention effects that confound measurement.
• Clinic-based structured tasks may outperform home naturalistic conversations for certain diagnostic classifications.
• Longitudinal voice features may correlate with mood or fatigue but not with clinically meaningful CNS outcomes.
• Models may fail to generalize across language, accent, education, or device conditions.
• High-frequency data may increase burden or privacy concern, reducing feasibility in real trials.

The strongest version of the hypothesis is therefore not that naturalistic home voice is universally better. The falsifiable claim is narrower: in specified CNS contexts of use, repeated home voice sampling may add clinically meaningful longitudinal information beyond clinic-only speech assessment.

Daily voice data collected through natural conversations at home may represent a valuable new data layer for CNS digital biomarker research. Its potential strength is not laboratory control, but ecological validity, repeated measurement, personal baseline estimation, and longitudinal sensitivity.

Clinic-based assessments remain essential. They provide standardization, diagnostic context, and clinical anchors. But they capture limited snapshots under artificial conditions. For CNS disorders where symptoms fluctuate and progression is gradual, daily home voice sampling may provide a richer view of real-world functioning.

The appropriate research path is hybrid and validation-driven: combine standardized clinic speech tasks with daily home conversations; define the context of use; validate technical reliability, algorithmic performance, clinical relevance, and usability; and treat early measures as exploratory until sufficient evidence is established.

If validated, naturalistic daily voice sampling could support decentralized trials, earlier detection of meaningful change, improved endpoint sensitivity, richer phenotyping, and more patient-centered CNS research. The opportunity is not to replace clinical assessment, but to measure what clinic visits often miss: how people speak, fluctuate, and function in daily life.