Skip to main content
AAO-HNSF 2023 Annual Meeting & OTO Experience Main banner
Icon Legend
This session is not in your schedule.
This session is in your schedule. Click again to remove it.
Back
Rate ePoster

Laryngology/Broncho-Esophagology

(440) Scoping Review of Voice Datasets Available for Voice AI Research

Monday, October 2, 2023
2:45 PM – 3:45 PM East Coast USA Time
Has Audio

    Presenting Author(s)

  • Ariel Farb, MS

    Data Manager
    Weill Cornell Medicine
    ny, New York, United States

    • Co-Author(s)

  • yael Bensoussan, MD

    Assistant Professor
    Department of Otolaryngology – Head & Neck Surgery, University of South Florida, Tampa, Florida, USA
    Tampa, FL, United States

  • Alisha Kodibagkar, MSE

    Technical Associate
    Massachusetts Institute of Technology
    Frisco, Texas, United States

  • SG

    Satrajit S. Ghosh, PhD

    Massachusetts Institute of Technology, USA

  • Anais Rameau, MD MPhil MS

    Assistant Professor
    Weill Cornell Medicine
    New York, New York, United States

    • Senior Author(s)

  • MP

    Maria E. Powell, PhD, CCC-SLP

    Assistant Professor of Otolaryngology-Head and Neck Surgery
    Vanderbilt University Medical Center
    Nashville, Tennessee, United States

Disclosure(s):
Ariel Farb, MS: No relevant relationships to disclose.
Alisha Kodibagkar, MSE: No relevant relationships to disclose.
Anais Rameau, MD MPhil MS: No relevant relationships to disclose.


Introduction: NIH’s Bridge2AI-Voice consortium aims to develop a large ethically sourced, diverse databank of human voice to fuel voice AI research. As part of a needs assessments for this project, the purpose of this study is to perform a scoping review of acoustic datasets currently available for voice AI research.

Methods: Pubmed, IEEE, and PMC databases were queried for voice AI articles published from 2000-2022. Studies were included if they were published in English and employed AI or machine learning approaches to classify acoustic data in the voice disordered populations. Publicly available and private datasets used in these articles were identified and reviewed. Additionally, we queried the Linguistics Data Consortium, GitHub, KBase, Mendeley Data, Figshare, and UCI machine learning repositories for speech samples from patient populations and healthy controls. Analysis focused on accessibility, diseases studied, content, and demographics.

Results: A total of 28 datasets were identified through the article review, and 169 datasets were identified through the listed repositories. Datasets identified in the repositories were cross referenced with article datasets, and a combined total of 187 unique datasets were analyzed. Most datasets (71.6%) were publicly available. Only 58.5% included diagnostic labels, of which 33.9% were related to voice disorders. Other diagnostic categories represented included neurological (43.5%), mood (11.3%), pediatric (9.7%), and respiratory (1.6%) disorders.

91.8% of datasets included sustained phonation, whereas fewer included standardized speech (36.1%) or free speech (1.6%) tasks. Languages represented included English, Arabic, French, Mandarin, German, Italian, Spanish, and Turkish. The majority reported sex (97.8%) and age (84.9%), but only 7.4% reported race and ethnicity.

Conclusions: Robust AI models require large and representative data from diverse populations, environments, and speech tasks. Although multiple acoustic datasets are currently available, existing data are lacking free speech tasks and clear documentation of a diverse population.