AxonVoice: An AI Tool for Voice-Based Digital Biomarker Phenotyping in Neurological Research

Principal Investigators

Avimanyu Bandyopadhyay, PhD
Chief Technology Officer
AxonDAO LLC
Email: avi@axondao.io
ORCID: 0000-0003-4806-2238

Tharusha Jayasena, PhD
Research Lead/Advisor
AxonDAO LLC
Email: tharusha@axondao.io
ORCID: 0000-0001-9655-9359

1. Overview

Scientific Rationale and Importance

Voice-based digital biomarkers offer a scalable, non-invasive method to phenotype neurological and psychological states, which is constrained by heterogeneous protocols and non-reproducible feature pipelines [1-4, 10, 11]. The AxonVoice Phase I project addresses this by developing a standardized, AI-ready infrastructure for voice data intake and phenotyping. This infrastructure-first approach is crucial for establishing reliable, reusable voice features essential for advancing biomedical research and enabling future AI-assisted analysis workflows.

Confirmation of Funding/Facilities/Ethics Approvals

This Phase I project requests funding as detailed in Section 5. All technical development will leverage existing AxonDAO technical infrastructure (e.g., secure servers, cloud services) and established open-source tools (e.g., openSMILE, Praat). This project is scoped as a non-clinical tool development and validation effort. Activities in Phase I will include documentation and protocol development to support a future Institutional Review Board (IRB) submission. Formal ethics review will be undertaken in later phases. All data collection procedures will follow informed, opt-in consent protocols.

Estimated Timeline

The project is scoped to be feasible within a 6-month timeframe.

Milestone	Target Completion (Month)
Finalize Intake Protocol & Architecture Specs	1
Implement Web-based Intake MVP & Anonymization	2
Implement & Validate Feature Extraction Pipeline	3
Commence Pilot Data Collection & Feature Stability Checks	4
Composite Score Development & Baseline Modeling	5
Analytical Validation, Preregistration & Reporting Draft	6

Data Sharing Commitment

Data sharing will adhere to open-science principles. Feature schemas, pipeline specifications, and documentation will be publicly released. Numerical feature vectors (Tier 3), which are non-identifiable and model-ready, are the primary output for responsible reuse. Raw audio (Tier 1) and optional transcripts (Tier 2) are high sensitivity and will be managed with strict access controls and participant deletion pathways. A preprint reporting the results will be submitted upon project completion.

2. Introduction

Theoretical Background, Key Literature, and Research Questions

Human voice carries acoustic and linguistic signatures associated with neurological and psychological states, including depression severity and cognitive decline [5-9]. Features like speech rate, pause structure, prosody, and semantic coherence are established correlates [1-9]. However, the field is hindered by a lack of standardization in intake protocols, non-transparent feature extraction, and insufficient analytical validation, leading to isolated, study-specific biomarker development [10, 11]. AI tools, while promising, require high-quality, structured data, which is currently lacking in the voice biomarker domain.

The overarching research question is: Can an AI-ready, standardized, and privacy-preserving infrastructure be developed to reliably phenotype voice-based digital biomarkers?

Numbered Hypotheses

H1 (Pipeline Robustness): The standardized web-based voice data intake pipeline will achieve a high completion and low failure rate (e.g., >95% successful audio file ingestion) across diverse participant devices, demonstrating technical robustness.

H2 (Feature Reproducibility): The defined acoustic and linguistic features extracted using the standardized, version-controlled pipeline will demonstrate high test-retest reliability and internal consistency across sessions and individuals, suggesting analytical stability.

H3 (Composite Score Stability): Variance component analysis using mixed-effects models will be conducted to partition within-subject and between-subject variance across repeated samples. Intraclass Correlation Coefficients (ICC) will be computed to quantify stability. Prediction: Within-subject variance across short-term repeated samples will be significantly lower than between-subject variance, indicating measurable individual-specific stability of the composite phenotype score over time.

Study Importance and Impact Statement

This project’s importance lies in its infrastructure-first approach. By prioritizing reproducible data systems over proprietary predictive model performance, AxonVoice addresses a foundational gap that limits cross-study comparability and independent validation. The successful development of this modular, privacy-aware phenotyping infrastructure will provide a critical tool for both academic and Decentralized Science (DeSci) researchers, accelerating the integration of voice as a reliable digital biomarker alongside other multi-modal datasets. This will enable more scientifically grounded and transparent AI agents in biomedical research, ultimately expanding the pool of validated, reusable data for global scientific inquiry, regardless of the initial results of the exploratory composite score.

3. Methods

Participants/Sample

Define Target Sample

The target sample consists of approximately 100-200 unique participants from the decentralized science (DeSci) community and research-interested individuals.

Inclusion/Exclusion Criteria

Inclusion Criteria: Adults (18+ years of age) who provide informed, opt-in consent and are able to record clear, short speech samples via a web-enabled device.

Exclusion Criteria: Inability to provide informed consent; known significant speech or language disorder that would preclude standard acoustic analysis.

Recruitment Strategy

Recruitment will utilize targeted outreach campaigns within the decentralized science and broader research community, leveraging community platforms and communication channels to promote informed, opt-in participation (supported by the budgeted funds for outreach).

Sample Size Justification (including power analysis)

An initial target of 100-200 unique voice samples is selected to balance feasibility with analytical robustness. This size provides sufficient diversity to characterize the distribution and stability of core acoustic and linguistic features across individuals. This is consistent with prior digital biomarker research focused on robust feature characterization (e.g., studies below 200 participants) [12, 13]. Specifically, similar prospective cohort studies [14] validated vocal biomarker scores using short recordings in cohorts around 100 participants, confirming the feasibility of short-sample voice phenotyping for feature characterization. The size is appropriate for Phase I objectives focused on infrastructure validation and feature reliability, not clinical inference, thus a formal power analysis for clinical prediction is not applicable.

Materials and Procedures

Experimental/Observational Procedures, Stimuli, Tasks, and Data Collection Settings

This is a tool-development and validation study using prospective, opt-in data collection.

Voice Data Intake: A standardized, web-based platform will be developed for data intake. Participants will access the platform and provide:

One or more 30–45 second speech samples (.wav format).
A simple self-reported mood or stress rating (1-10 scale).
A one-time demographic survey (age range, sex, education level, country).

A subset of participants will be invited to provide repeated recordings across short-term intervals (e.g., 2-4 recordings within a 2-4 week window) to enable test-retest reliability and longitudinal stability analyses. Participation in repeat recordings will remain voluntary but will be encouraged to ensure adequate within-subject variance modeling for Hypotheses 2 and 3.

Guidance: Participants will receive brief, standardized guidance (e.g., quiet environment, typical speaking voice) to reduce device-related acoustic variability while maintaining ecological validity.

Stimuli/Tasks: The task will involve reading a short, standardized text or responding to a brief, open-ended prompt designed to elicit 30-45 seconds of speech.

Equipment: Participants' personal web-enabled devices (smartphones, computers with microphones).

Planned Analyses

Outline Analysis Pipeline and Preprocessing Steps

Voice recordings will undergo an established, standardized, and fully documented preprocessing pipeline using open-source signal processing libraries (openSMILE, Praat, librosa, pyAudioAnalysis). Steps will include:

Audio Quality Check: Ensure minimum duration, acceptable signal-to-noise ratio.
Normalization: Volume and time-alignment normalization.
Feature Extraction: Extraction of core acoustic measures (speech rate, pause ratio, pitch, intensity, jitter) and optional linguistic features (lexical diversity, semantic coherence, sentiment). All steps will be reproducible, version-controlled, and transparently documented.

Statistical Tests and Hypothesis-Specific Predictions

Hypothesis	Statistical Test	Prediction
H1 (Pipeline Robustness)	Descriptive statistics (success rate, failure logs).	>95% success rate for ingestion across devices.
H2 (Feature Reproducibility)	Intraclass Correlation Coefficients (ICC) for test-retest reliability; distributional analysis.	ICC > 0.7 for core acoustic features (e.g., speech rate, pitch variability).
H3 (Composite Score Stability)	Mixed-effects variance component modeling; ICC computation	Within-subject variance will be significantly lower than between-subject variance.

Contingent Analysis Decisions

IF any core acoustic features (e.g., pitch variability) show poor test-retest reliability (ICC < 0.5), THEN the feature will be flagged as unstable for future use, and the feature extraction parameters will be re-optimized before final pipeline validation.

IF device-related acoustic artifacts significantly skew feature distributions, THEN additional preprocessing steps (e.g., noise reduction filters) will be integrated and documented, and the affected samples will be reprocessed.

Ethics and Data Management

Ethical Approval Status or Plan

This Phase I project is a non-clinical, wellness-oriented tool development effort. All participants will provide explicit, opt-in consent. Preparatory activities for a future IRB submission (e.g., protocol documentation, consent materials) are included in Phase I; however, the formal submission and review process fall outside the scope of this phase. No diagnostic or clinical interpretation will be provided to participants. Composite scores are exploratory and not health advice.

Confidentiality Measures

A three-tier data architecture supports confidentiality: (1) Raw Audio Recordings (Tier 1: Highest sensitivity, stored securely, deletable upon request); (2) Transcripts (Tier 2: Optional, lower sensitivity); and (3) Numerical Feature Vectors (Tier 3: Non-identifiable, model-ready, primary output). Access controls, consent flows, and deletion pathways will be explicitly documented. All data will be encrypted in transit using TLS (HTTPS) protocols and encrypted at rest using industry-standard encryption (e.g., AES-256). Secure cloud infrastructure with role-based access control will be used to restrict access to Tier 1 (raw audio) and Tier 2 (transcript) data.

Data storage jurisdiction will comply with applicable data protection standards and will be transparently documented prior to participant intake. Raw audio files (Tier 1) will have a defined retention policy (e.g., retained only for the duration necessary for feature extraction and validation unless explicit participant consent is provided for extended archival). Participants will have a documented pathway to request deletion of Tier 1 and Tier 2 data at any time.

Tier 3 numerical feature vectors, which are non-identifiable and stripped of raw acoustic reconstruction capability, will constitute the primary long-term archival output.

Data Sharing/Archiving Plan

The study protocol will be preregistered on ResearchHub. Data sharing will follow open-science principles. Feature schemas, specifications, pipeline documentation, and the final feature set (Tier 3 data) will be archived and released publicly to support independent research and validation.

4. Pilot Data

Not applicable. The project relies on existing findings in the voice biomarker literature [1-14] to justify the methods and sample size, focusing on infrastructure validation rather than preliminary clinical findings.

5. Budget

This budget covers costs essential for Phase I infrastructure validation, excluding the core front-end/back-end development supported by AxonDAO.

Item	Estimated Cost	Rationale
Participant recruitment and research outreach	$1,000	Targeted outreach (platforms, incentives) for 100-200 unique participants.
Project execution and validation costs	$2,500	Protocol specification, analytical validation, QC analysis, documentation, preregistration, and reporting.
Compute, storage, and tooling	$1,500	Batch audio processing, feature extraction, temporary storage, and compute for testing/validation.
Total Request	$5,000

Core engineering and front-end/back-end development costs are supported by AxonDAO internal infrastructure and personnel resources. The requested funds are strictly allocated to research execution, analytical validation, participant recruitment, compute/testing, documentation, preregistration, and reporting activities required for Phase I infrastructure validation.

6. References

Low DM, Bentley KH, Ghosh SS. Automated assessment of psychiatric disorders using speech: A systematic review. Laryngoscope Investig Otolaryngol. 2020 Jan 31;5(1):96-116.
Zhao Q, Fan HZ, Li YL, Liu L, Wu YX, Zhao YL, Tian ZX, Wang ZR, Tan YL, Tan SP. Vocal Acoustic Features as Potential Biomarkers for Identifying/Diagnosing Depression: A Cross-Sectional Study. Front Psychiatry. 2022 Apr 28;13:815678.
Schewski L, Doss MM, Beldi G, Keller S. Measuring negative emotions and stress through acoustic correlates in speech: A systematic review. PLoS One. 2025 Jul 24;20(7):e0328833.
Menne F, Lindsay H, Tröger J, Paulmann S, König A, Steinbach N, Reif A, Plichta MM, Schmidt-Kassow M. Voice as objective biomarker of stress: association of speech features and cortisol. Acta Neuropsychiatr. 2025 Sep 3;37:e84.
Wiseman M, Yep R, Wood Alexander M, Pople CB, Perri L, Gopinath G, Vasileiadi M, Robin J, Spilka MJ, Simpson W, Yunusova Y, Munoz DP, Coe BC, Brien D, Nestor S, Lipsman N, Giacobbe P, Rabin JS. Objective speech measures capture depressive symptoms and associated cognitive difficulties. Transl Psychiatry. 2025 Nov 17;15(1):525.
Menne F, Dörr F, Schräder J, Tröger J, Habel U, König A, Wagels L. The voice of depression: speech features as biomarkers for major depressive disorder. BMC Psychiatry. 2024 Nov 12;24(1):794.
Koops S, Brederoo SG, de Boer JN, Nadema FG, Voppel AE, Sommer IE. Speech as a Biomarker for Depression. CNS Neurol Disord Drug Targets. 2023;22(2):152-160.
Cohen AS, Divers R, Calamia M, Masucci M, Granrud OE, Corporandy A, Warren KK, Elvevåg B, Chandler C, Diaz-Asper C. Speech pause and speech rate for evaluating Alzheimer's and mild cognitive impairment: A meta-analysis. J Int Neuropsychol Soc. 2025 Dec 11:1-8.
Fagherazzi G, Fischer A, Ismael M, Despotovic V. Voice for Health: The Use of Vocal Biomarkers from Research to Clinical Practice. Digit Biomark. 2021 Apr 16;5(1):78-88.
Robin J, Harrison JE, Kaufman LD, Rudzicz F, Simpson W, Yancheva M. Evaluation of Speech-Based Digital Biomarkers: Review and Recommendations. Digit Biomark. 2020 Oct 19;4(3):99-108.
Briganti G, Lechien JR. Speech and Voice Quality as Digital Biomarkers in Depression: A Systematic Review. J Voice. 2025 May 22:S0892-1997(25)00187-0.
von Polier GG, Ahlers E, Volkening J, Langner J, Patil KR, Eickhoff SB, Helmhold F, Krautz AE, Langner D. Exploring voice as a digital phenotype in adults with ADHD. Sci Rep. 2025 May 24;15(1):18076.
Riad R, Denais M, de Gennes M, Lesage A, Oustric V, Cao X, Mouchabac S, Bourla A. Automated Speech Analysis for Risk Detection of Depression, Anxiety, Insomnia, and Fatigue: Algorithm Development and Validation Study. J Med Internet Res. 2024;26:e58572.
Larsen E, Murton O, Song X, Joachim D, Watts D, Kapczinski F, Venesky L, Hurowitz G. Validating the efficacy and value proposition of mental fitness vocal biomarkers in a psychiatric population: prospective cohort study. Front Psychiatry. 2024 Mar 5;15:1342835.

Funding

Funding

AxonVoice: An AI Tool for Voice-Based Digital Biomarker Phenotyping in Neurological Research

Funding Progress

Nonprofit Recipient

Topics

DOI

Funding Progress

Nonprofit Recipient

Topics

DOI

AxonVoice: An AI Tool for Voice-Based Digital Biomarker Phenotyping in Neurological Research

Funding Progress

Nonprofit Recipient

Topics

DOI