Bridge2AI-Voice v2.0: Clinical Voice Biomarker Dataset

• Bridge2AI-Voice v2.0 is a multifaceted clinical voice dataset, ethically sourced from 442 participants across five sites, with comorbid diagnostic labels.
• It employs a BIDS-aligned schema with detailed metadata, enabling privacy-preserving, AI-ready workflows for diverse classification and screening tasks.
• Methodological studies use domain-adaptive self-supervised learning and multi-task architectures to achieve improved Macro F1 scores and robust acoustic feature representations.

Searching arXiv for the cited Bridge2AI-Voice papers to ground the article in current preprints. I’ll look up the main arXiv record and closely related Bridge2AI-Voice papers before drafting the entry. Bridge2AI-Voice v2.0 is a multi-institutional clinical voice dataset developed within the NIH Bridge2AI initiative for using human voice as a biomarker of health. In the cited methodological studies, it is described as an ethically sourced, diverse voice dataset linked to health information, distributed through a Bridge2AI-Voice v2.0.0 release on PhysioNet and organized with a BIDS-aligned schema for AI/ML use (Liu et al., 29 Jan 2026). Within the broader Bridge2AI program, it serves as infrastructure for screening, diagnosis, treatment support, and voice AI research, while also functioning as a testbed for privacy-conscious feature representations, metadata-rich packaging, and clinically grounded learning paradigms (Caufield et al., 12 Sep 2025).

1. Programmatic setting within Bridge2AI

Bridge2AI-Voice v2.0 is associated with the Precision Public Health Grand Challenge, also referred to as “Voice as a Biomarker of Health.” The stated objective of that project is to use human voice recordings to identify possible links to disease states and to develop computational models for screening, diagnosis, treatment support, and voice AI research (Caufield et al., 12 Sep 2025).

In Bridge2AI terms, AI-readiness is not treated as a property of data availability alone. The consortium defines AI-readiness through criteria including FAIRness, provenance, characterization, pre-model explainability, ethics, sustainability, and computability, and the voice project is presented as one of the Grand Challenges in which those requirements must be operationalized through metadata, packaging, and governance (Caufield et al., 12 Sep 2025). This places Bridge2AI-Voice v2.0 in a distinct category relative to smaller task-specific speech corpora: it is intended not merely to support a single benchmark, but to support reusable biomedical AI workflows under explicit data-management constraints.

The published studies also emphasize that voice is being treated as a scalable, non-invasive biomedical signal. That framing appears in both downstream classification work and metadata-design work, where the dataset is positioned as a resource for clinically meaningful, privacy-aware modeling rather than as a generic speech collection (Piao et al., 28 Aug 2025).

2. Cohort composition and clinical label structure

The cited studies report a cohort of 442 unique participants recruited across five North American sites or institutions (Liu et al., 29 Jan 2026). One methodological study reports that, after cleaning and preprocessing, the dataset comprised 16,738 distinct recordings (Liu et al., 29 Jan 2026). A separate study using a curated subset of the same release reports over 19,000 recordings from 442 participants across five North American clinical institutions (Piao et al., 28 Aug 2025). The difference reflects study-specific preprocessing and subsetting rather than a contradictory cohort identity.

For multi-label disease classification, diagnoses are grouped into four broad categories, with comorbidity explicitly allowed because a single participant may have multiple disorders (Liu et al., 29 Jan 2026).

Broad category	Participants
Voice disorders	230 (52.0%)
Respiratory disorders	226 (51.1%)
Neurological / neurodegenerative disorders	160 (36.2%)
Mood / psychiatric disorders	145 (32.8%)

These percentages do not sum to 100% because the dataset is used in a comorbid, multi-label setting (Liu et al., 29 Jan 2026). That feature is methodologically important: it makes the corpus suitable for formulations in which disease categories are not mutually exclusive, which is closer to real clinical populations than single-diagnosis benchmarks.

A second line of work on the same dataset defines nine binary classification tasks spanning respiratory, voice, and neurological disorders: Airway Stenosis, Asthma, Chronic Obstructive Pulmonary Disease (COPD), Laryngeal Cancer, Benign Lesions of the Vocal Cord, Spasmodic Dysphonia / Laryngeal Tremor, Vocal Fold Paralysis, Parkinson’s Disease, and Alzheimer’s Disease / Mild Cognitive Impairment (AD/MCI) (Piao et al., 28 Aug 2025). This demonstrates that Bridge2AI-Voice v2.0 supports both broad disease-family labeling and more specific condition-level prediction tasks.

3. Data organization, metadata, and governance

The voice Grand Challenge uses a BIDS-aligned organization schema. The reported file ecosystem includes WAV for audio, TSV for clinical and phenotypic form data, JSON sidecar files for key metadata, JSON data dictionaries, and Derivatives for AI/ML-ready interaction (Caufield et al., 12 Sep 2025). The dataset is also linked to a “Voice as a Biomarker for AI Health” FHIR R4 profile, derived from the US Core STU5 profile, for interoperability with broader health-data systems (Caufield et al., 12 Sep 2025).

The metadata pipeline is described as follows: raw audio and questionnaire data are first stored in REDCap, then converted into BIDS structure; data dictionaries are generated from ReproSchema-structured protocols and the SenseLab audio feature extraction toolkit, and those dictionaries are stored in BIDS JSON format (Caufield et al., 12 Sep 2025). This makes the corpus not only machine-readable but also protocol-linked, which is important for reproducibility and auditability.

Metadata in the voice project connect recordings to demographics, habits with health impact such as smoking and drinking, and disease-specific questionnaires including Voice Handicap Index-10 (VHI-10), Patient Health Questionnaire (PHQ-9), and General Anxiety Disorder (GAD-7) (Caufield et al., 12 Sep 2025). The project also distinguishes between a fixed feature format, in which static features are extracted from the entire waveform, and a temporal format, in which features vary with the length of each recording (Caufield et al., 12 Sep 2025). This supports both tabular ML and sequence-aware modeling.

Governance and privacy are treated as first-order design constraints. The voice Grand Challenge is reported to have made an initial release on Health Data Nexus without raw audio waveforms or personally identifying information, alongside a separate release through PhysioNet (Caufield et al., 12 Sep 2025). The same report notes that the team had to develop novel anonymization strategies because there were few consistent standards for collecting voice data, regardless of privacy-preservation strategies (Caufield et al., 12 Sep 2025). That point is central to the status of Bridge2AI-Voice v2.0: the corpus is both a dataset and a standards-development case for identifiable biomedical audio.

4. Acoustic representations and feature modalities

Bridge2AI-Voice v2.0 has been used with both raw-audio-derived spectrotemporal representations and privacy-preserving derived features. In one multi-task study, the public version is described as providing pre-extracted acoustic features such as MFCCs, log-Mel spectrograms, and handcrafted voice descriptors, while excluding raw audio (Piao et al., 28 Aug 2025). This feature-only availability is tightly aligned with a privacy-by-design modeling strategy because it avoids raw audio transmission.

In the domain-adaptive self-supervised study, the deep input is a log-mel spectrogram computed from audio sampled at 16 kHz using a 400-point FFT, 25 ms window, 10 ms hop, 128 mel bands, and a log-decibel scale, yielding a representation $\mathbf{S} \in \mathbb{R}^{128 \times T}$ (Liu et al., 29 Jan 2026). The same work also uses 131 handcrafted/static acoustic features, including jitter, shimmer, and HNR-type clinical acoustic features from OpenSMILE, together with pitch, formants, and prosodic features from Praat / Parselmouth (Liu et al., 29 Jan 2026).

These two representational regimes—time-frequency tensors for deep encoders and static descriptors for clinically familiar acoustic measurement—are not treated as mutually exclusive. Instead, they are combined in downstream models by concatenating learned deep features with the 131-dimensional static vector (Liu et al., 29 Jan 2026). This design choice is notable because it places Bridge2AI-Voice v2.0 at the intersection of conventional clinical phonetics and modern representation learning, rather than substituting one for the other.

5. Domain-adaptive self-supervised learning on Bridge2AI-Voice v2.0

A major methodological use of Bridge2AI-Voice v2.0 is the study of domain-adaptive self-supervised learning for clinical voice-based disease classification. The problem is framed as multi-label disease classification because participants may have multiple comorbid conditions, and the authors ask whether a standard Masked Autoencoder (MAE) configuration can be better adapted to in-domain clinical spectrograms by optimizing reconstruction loss, normalization, and masking (Liu et al., 29 Jan 2026).

The model uses a two-stage framework. In stage 1, an asymmetric MAE encoder-decoder is built on the Audio Spectrogram Transformer (AST): the spectrogram is split into non-overlapping patches, only visible patches are fed to the AST encoder, and a lightweight Transformer decoder reconstructs the missing patches from encoded visible tokens plus learnable $[MASK]$ tokens (Liu et al., 29 Jan 2026). In stage 2, the pretrained AST encoder produces a 768-dimensional deep feature vector, which is concatenated with the 131-dimensional static feature vector and fed to an Attention-based Feed-Forward Neural Network (Attention-FFNN) for final multi-label prediction (Liu et al., 29 Jan 2026).

The study systematically compares Mean Absolute Error (MA-Error) versus Mean Squared Error, patch-wise normalization versus no/global normalization, and content-aware masking versus random masking. The interpretation given is that MSE over-penalizes large deviations and tends to emphasize high-energy components such as stable formants, whereas MA-Error is less sensitive to outliers and gives more balanced attention to both dominant and subtle spectral cues. That distinction is presented as especially relevant for pathological voice, where diagnostic information may lie in low-energy irregularities such as breathiness, turbulence, aperiodic bursts, jitter, shimmer, and harshness (Liu et al., 29 Jan 2026).

The best-performing configuration combines MA-Error loss, patch-wise normalization, and content-aware masking, reaching Macro F1 = $0.688 \pm 0.009$ over 10 downstream fine-tuning runs. The strongest out-of-domain SSL baseline, SSAST pretrained on AudioSet, reports Macro F1 = $0.663 \pm 0.011$, so the optimized in-domain model improves Macro F1 by 0.025 absolute (Liu et al., 29 Jan 2026). The ablation study reports 0.608 for MA-Error (Base), 0.655 for MA-Error + CA, and 0.688 for MA-Error + Norm + CA, while the best MSE-based variant, MSE + Norm + CA, reaches 0.641 (Liu et al., 29 Jan 2026).

The paper’s explicit conclusion is that standard MAE settings are not optimal for pathological voice and that objective-level adaptation matters more than simply using larger out-of-domain pretraining corpora (Liu et al., 29 Jan 2026). In relation to Bridge2AI-Voice v2.0, this establishes the dataset as more than a downstream benchmark: it is the in-domain substrate that makes domain-adaptive SSL itself empirically testable.

6. Unified multi-task clinical detection

A second major research use of Bridge2AI-Voice v2.0 is MARVEL (“Multi-task Acoustic Representations for Voice-based Health Analysis”), a dual-branch, multi-task learning framework that simultaneously detects nine distinct neurological, respiratory, and voice disorders using only derived acoustic features (Piao et al., 28 Aug 2025). The study is explicitly motivated by the observation that existing approaches often focus on single conditions and do not exploit the multi-faceted information embedded in speech.

MARVEL processes two acoustic views: MFCCs and log-Mel spectrograms. It uses ResNet18 for the MFCC branch and EfficientNet-B0 for the spectrogram branch, concatenates the resulting modality embeddings into a 1792-dimensional fused representation, reduces that representation to a shared 512-dimensional latent backbone, and attaches task-specific two-layer MLP heads for the nine binary tasks (Piao et al., 28 Aug 2025). The model is trained with weighted binary cross-entropy and balanced batches containing 6 positive and 6 negative samples per task, for 9 tasks $\times$ 12 = 108 samples per batch (Piao et al., 28 Aug 2025).

Reported performance includes overall AUROC = 0.78, grouped AUROC values of 0.89 for neurological disorders, 0.76 for voice disorders, and 0.74 for respiratory disorders, and per-task AUROCs including 0.89 for Airway Stenosis, 0.79 for COPD, 0.87 for Vocal Fold Paralysis, 0.81 for Parkinson’s Disease, and 0.97 for AD/MCI (Piao et al., 28 Aug 2025). The authors state that MARVEL consistently outperforms single-modal baselines by 5–19% and surpasses self-supervised baselines on 7 of 9 tasks (Piao et al., 28 Aug 2025).

The study also reports a correlation analysis between MARVEL’s learned embeddings and the same family of 131 handcrafted acoustic features from openSMILE and Praat, finding condition-specific alignment with clinically meaningful descriptors such as HNR and formant-related features for laryngeal cancer, glottal-source irregularity features for benign vocal cord lesions, MFCC-related spectral envelope features for COPD, loudness and intensity instability for vocal fold paralysis, and prosodic/slope-related measures for Parkinson’s disease (Piao et al., 28 Aug 2025). This is important because it links data-driven internal representations back to recognized acoustic biomarkers rather than treating the model as an uninterpretable black box.

7. Open issues, standardization gaps, and broader significance

The published metadata analysis is explicit that there are few consistent standards for collecting voice data, and that privacy-preserving strategies for voice are not standardized (Caufield et al., 12 Sep 2025). It also states that novel anonymization strategies had to be developed within the voice Grand Challenge and that future releases will implement governance principles for access to audio waveforms (Caufield et al., 12 Sep 2025). These are not ancillary details: they identify the principal unresolved problem for Bridge2AI-Voice v2.0 as the joint management of scientific utility, interoperability, and re-identification risk.

The broader Bridge2AI standards work argues that AI-ready datasets are FAIR, fully reliable, robustly defined, and computationally accessible, and that provenance, quality control, file-level manifests, standard schemas, and controlled vocabularies are necessary to prevent AI systems from learning technical artifacts instead of domain signal (Cannon et al., 12 Dec 2025). A plausible implication for Bridge2AI-Voice v2.0 is that voice metadata must be detailed enough to control for recording device effects, environment and channel effects, preprocessing history, and annotation provenance, in the same way that sequencing center effects and pipeline differences must be controlled in genomics.

Taken together, the cited work presents Bridge2AI-Voice v2.0 as a biomedical voice resource with three simultaneous identities. It is a clinically grounded cohort with comorbid labels and multi-site collection; it is a metadata-rich, BIDS-aligned and privacy-conscious data product; and it is a methodological platform on which both domain-adaptive self-supervision and unified multi-task learning have been evaluated at scale (Liu et al., 29 Jan 2026). This suggests that its significance lies not only in the number of recordings or diagnoses, but in the combination of dataset design, governance, and representational flexibility that enables research on clinically meaningful voice biomarkers under AI-readiness constraints.