Objective:

The introduction motivates automated voice-disorder classification by describing the diversity of etiologies and overlapping acoustic features, and notes the potential of AI-based assessment of brief voice recordings.

Approach:

• Clinical background: Voice disorders are classified as functional, organic, or neurological, and acoustic features can overlap across these etiological groups.
• Prior research: The introduction describes high sensitivity and specificity reported by prior AI studies using the VOICED database, which contains 208 samples across 4 diagnostic classes, and contrasts it with the more diverse Saarbrücken Voice Database, which contains 2225 speaker or recording-session entries across 72 diagnostic categories.
• Study method: The supplied excerpt ends during the introduction and does not describe the current study’s models, data-processing methods, or evaluation design.

Key Findings:

• Prior VOICED-based work reported sensitivity of 0.996 and specificity of 0.989, but the dataset contains only 208 samples.
• On the SVD, prior studies reported sensitivity and specificity of 0.82 and 0.83, respectively, for regression-based methods.
• An SVD study using an ensemble of transformer models reported accuracy of 0.91 (SD 0.006) and area under the curve of 0.91 (SD 0.005).

Interpretation:

The introduction identifies overlapping acoustic features and the limited size and diagnostic range of the VOICED database as challenges for automated voice-disorder classification and generalizability.

Limitations:

• The supplied excerpt ends during the introduction and does not include the current study’s methods or results.

Sources:

• Peng et al. (2023), cited in the source for VOICED sensitivity and specificity.
• Cesari et al. (2018), cited in the source for the VOICED database.
• Pützer and Koreman (1997), cited in the source for the Saarbrücken Voice Database.