=== YAML Fixing Applied ===
id: bridge2ai-voice-v1-0
name: Bridge2AI-Voice
title: "Bridge2AI-Voice: An ethically-sourced, diverse voice dataset linked to health information (v1.0)"
description: >
  Bridge2AI-Voice is a comprehensive, ethically sourced dataset linking derived data from voice
  recordings to corresponding clinical information to enable research on voice as a biomarker of
  health. Version 1.0 provides 12,523 recordings for 306 adult participants collected across five
  sites in North America, focused on conditions known to manifest in voice (voice disorders,
  neurological and neurodegenerative disorders, mood and psychiatric disorders, and respiratory
  disorders). This initial release is designed to be low risk: original audio waveforms are
  omitted; only derived artifacts (e.g., spectrograms, acoustic/phonetic/prosodic features) and
  de-identified phenotype data are provided. Data collection followed a standardized protocol
  (demographics, validated questionnaires, targeted confounders, and structured voice tasks) using
  a custom tablet application and, when possible, a headset microphone. Raw audio was standardized
  (mono, 16 kHz with a Butterworth anti-aliasing filter) prior to derivations. Extensive HIPAA Safe
  Harbor de-identification was applied, transcripts of free speech were removed, state/province
  suppressed, and only country retained. Detailed data dictionaries accompany feature and phenotype
  files. For documentation, see https://docs.b2ai-voice.org/.
language: en
issued: 2024-11-27
version: "1.0"
doi: "doi:10.57764/qb6h-em84"
page: "https://doi.org/10.57764/qb6h-em84"
keywords:
  - voice
  - bridge2ai
  - audio
  - health
  - spectrograms
  - clinical
  - phenotype
created_by:
  - Alistair Johnson
  - Jean-Christophe Bélisle-Pipon
  - David Dorr
  - Satrajit Ghosh
  - Philip Payne
  - Maria Powell
  - Anaïs Rameau
  - Vardit Ravitsky
  - Alexandros Sigaras
  - Olivier Elemento
  - Yael Bensoussan
license: Bridge2AI Voice Registered Access License
purposes:
  - response: >
      Create an ethically sourced flagship dataset to enable AI research on voice as a biomarker
      of health and support insights into links between voice and clinical conditions.
tasks:
  - response: >
      Development and evaluation of AI methods to extract prognostically useful information from
      voice-derived data (e.g., spectrograms, acoustic/phonetic/prosodic features).
  - response: >
      Disease-related voice analytics across voice, neurological, mood/psychiatric, and
      respiratory disorders.
addressing_gaps:
  - response: >
      Address the need for a large, high-quality, multi-institutional, diverse voice dataset
      linked to other health information, collected with standardized protocols and ethical
      oversight.
funders:
  - grantor:
      name: National Institutes of Health (NIH)
    grant:
      name: "Bridge2AI: Voice as a Biomarker of Health - Building an ethically sourced, bioacoustic database to understand disease like never before"
      grant_number: 3OT2OD032720-01S1
subsets:
  - name: Adult cohort (v1.0)
    description: >
      v1.0 includes only adult participants. Disease-focused recruitment across five North
      American sites.
    is_data_split: "no"
    is_subpopulation: "yes"
  - name: spectrograms.parquet
    path: spectrograms.parquet
    media_type: application/x-parquet
    description: >
      Dense time-frequency representations (513 x N spectrograms) per recording; includes
      participant_id, session_id, task_name.
  - name: phenotype.tsv
    path: phenotype.tsv
    media_type: text/tab-separated-values
    description: >
      One row per participant with demographics, acoustic confounders, and validated questionnaire
      responses.
  - name: phenotype.json
    path: phenotype.json
    media_type: application/json
    description: Data dictionary for phenotype.tsv (one-sentence descriptions per column).
  - name: static_features.tsv
    path: static_features.tsv
    media_type: text/tab-separated-values
    description: >
      One row per recording with features derived from openSMILE, Praat/Parselmouth, and
      torchaudio.
  - name: static_features.json
    path: static_features.json
    media_type: application/json
    description: Data dictionary for static_features.tsv (feature descriptions).
instances:
  - representation: >
      Voice-derived artifacts (spectrograms, acoustic/phonetic/prosodic features) linked to
      de-identified phenotypic and clinical variables; per-participant and per-recording data.
    instance_type: >
      Participants, sessions, and recordings; per-recording feature rows; per-participant phenotype rows.
    data_type: >
      Derived data only in v1.0 (no original audio): spectrogram arrays, acoustic/phonetic/prosodic
      features; tabular phenotype and feature files with accompanying JSON data dictionaries.
    counts: 12523
    label: >
      No explicit task labels are provided; dataset includes condition cohorts and questionnaire data.
    sampling_strategies:
      - strategies: >
          Disease-focused recruitment at specialty clinics; participants pre-screened for inclusion/exclusion in
          five predetermined groups (respiratory, voice, neurological, mood/psychiatric, pediatric [not included in v1.0]).
        is_sample: "yes"
        is_random: "no"
        source_data: "Specialty clinics across five sites in North America"
        is_representative: "no"
        why_not_representative: >
          Targeted recruitment by disease cohorts to cover conditions with voice manifestations.
    missing_information:
      - missing: Original audio waveforms
        why_missing: Omitted in v1.0 to reduce re-identification risk; planned for future releases with added safeguards
      - missing: Free speech transcripts
        why_missing: Removed during de-identification
relationships:
  - description:
      - Participants can have multiple sessions; sessions contain multiple recordings; identifiers include participant_id and session_id.
subpopulations:
  - identification:
      - Adult cohort only (v1.0)
      - Disease categories: voice disorders; neurological/neurodegenerative disorders; mood/psychiatric disorders; respiratory disorders
    distribution:
      - 306 participants across five North American sites (v1.0)
sensitive_elements:
  - description:
      - De-identified health-related phenotype and clinical information are included.
is_deidentified:
  description:
    - HIPAA Safe Harbor identifiers removed (e.g., names, specific dates finer than year, contact numbers, MRNs, full-face photos, etc.).
    - State and province removed; country of data collection retained.
    - Free speech transcripts removed.
    - Original audio waveforms omitted; only derived data released in v1.0.
acquisition_methods:
  - description:
      - Data collected directly from patients at specialty clinics via standardized protocol and structured voice tasks.
    was_directly_observed: "yes (raw audio recordings of voice tasks)"
    was_reported_by_subjects: "yes (validated questionnaires and demographics)"
    was_inferred_derived: "yes (spectrograms; acoustic/phonetic/prosodic features; automatic transcripts generated but not released)"
    was_validated_verified: >
      Protocols standardized; ethical review approvals obtained; preprocessing and feature extraction pipelines documented.
collection_mechanisms:
  - description:
      - Custom tablet application for data entry and recording; headset microphone used when possible; standardized clinical protocol.
data_collectors:
  - description:
      - Project investigators at specialty clinics and institutions across five North American sites.
ethical_reviews:
  - description:
      - Data collection and sharing approved by the University of South Florida Institutional Review Board.
      - Submission for review to the University of Toronto Research Ethics Board.
preprocessing_strategies:
  - description:
      - Audio converted to mono and resampled to 16 kHz with a Butterworth anti-aliasing filter.
      - Spectrograms computed via short-time FFT (25 ms window, 10 ms hop, 512-point FFT).
      - Acoustic features extracted using openSMILE.
      - Phonetic and prosodic features computed using Parselmouth and Praat (e.g., F0, formants, voice quality).
      - Automatic transcriptions generated using OpenAI Whisper Large (not included in v1.0 release).
    used_software:
      - name: openSMILE
        url: "https://audeering.github.io/opensmile/"
      - name: Praat
        url: "https://www.fon.hum.uva.nl/praat/"
      - name: Parselmouth
        url: "https://parselmouth.readthedocs.io/"
      - name: Torchaudio
        url: "https://pytorch.org/audio/"
      - name: OpenAI Whisper (Large)
        url: "https://github.com/openai/whisper"
labeling_strategies:
  - description:
      - Automatic speech transcriptions generated using OpenAI Whisper Large; transcripts were removed for this release.
raw_sources:
  - description:
      - Raw voice audio was collected but is not distributed in v1.0; future releases aim to include voice data with additional security measures.
future_use_impacts:
  - description:
      - Disease-focused recruitment may limit generalizability to broader populations; derived-only release reduces privacy risks but restricts certain analyses dependent on raw audio.
external_resources:
  - external_resources:
      - Project documentation website: https://docs.b2ai-voice.org
      - REDCap export/conversion tooling reference (Zenodo): https://doi.org/10.5281/zenodo.14148755
      - Preprocessing code (b2aiprep): https://github.com/sensein/b2aiprep
    archival:
      - Versioned DOI: https://doi.org/10.57764/qb6h-em84 (v1.0)
      - Latest DOI: https://doi.org/10.57764/3sg0-7440
distribution_formats:
  - description:
      - Parquet: spectrograms.parquet
      - TSV (tab-delimited): phenotype.tsv, static_features.tsv
      - JSON: phenotype.json, static_features.json
distribution_dates:
  - description:
      - Initial public release (credentialed access): 2024-11-27 (v1.0)
license_and_use_terms:
  description:
    - License: Bridge2AI Voice Registered Access License.
    - Access restricted to credentialed users who sign the Bridge2AI Voice Registered Access Agreement (DUA).
    - Required training: TCPS 2: CORE 2022.
    - Files are accessible only after meeting credentialing, training, and DUA requirements via Health Data Nexus.
maintainers:
  - description:
      - Health Data Nexus; Bridge2AI-Voice team. Corresponding author contact available to logged-in users on the resource page.
updates:
  description:
    - b2ai-voice v1.0 is the first release; future releases aim to include original voice audio with additional safeguards.
version_access:
  description:
    - Versioned DOI for v1.0: https://doi.org/10.57764/qb6h-em84
    - Latest version DOI: https://doi.org/10.57764/3sg0-7440
is_tabular: mixed (tabular TSV/JSON and array-based Parquet)