Benign and malignant lesions of the vocal folds can significantly impact voice quality and may cause serious health complications, including mortality in malignant cases. Often, the diagnosis of vocal fold pathology relies on specialized visualization techniques that may not always be accessible. The Bridge2AI-Voice dataset, funded by the NIH, aims to curate a diverse collection of ethically sourced voice recordings for AI research. Through initial efforts with the Bridge2AI Voice project, we hope to develop early screening tools to improve access to care, particularly for historically underrepresented populations.
The Bridge2AI-Voice v1.0 data set was used in this study. It consisted of a total of 12,523 audio recordings from 306 subjects from five sites in North America. Two groups of analysis were created. Group 1 compared laryngeal cancer (n = 10), benign cord lesions (CL) (n = 13), and no voice disorder (NVD) (n = 122). Group 2 included laryngeal cancer and no other voice disorder (NOVD) (n = 6), benign CL NOVD (n = 11), spasmodic dysphonia, no lesions (n = 8), and unilateral vocal fold paralysis (UVFP), no lesions (n = 26).
Acoustic features, including mean harmonic-to-noise ratio (HNR), HNR standard deviation (SD), mean local jitter, mean local shimmer, and mean fundamental frequency (F0), were extracted from the standardized “Rainbow Passage” using the openSMILE software. The statistical analysis involved the Kruskal-Wallis test, followed by Dunn’s post-hoc test with Holm’s correction for multiple comparisons. The analyses were also stratified by sex to account for physiological differences in vocal parameters.
No statistically significant differences were observed between the lesion-present (n = 23) and lesion-absent (n = 153) groups in terms of age, weight, gender identity, sexual orientation, race, or ethnicity. However, the lesion-present group included 12.8% more African Americans and had a mean weight approximately 20 pounds higher.
There were significant differences found in the mean HNR (p = 0.019), HNR SD (p = 0.028), and F0 (p = 0.012) across all participants in Group 1 between benign CL and NVD. Benign CL also differed from laryngeal cancer in HNR SD (p = 0.028). There were no significant differences in jitter or shimmer, and no significant differences were found in Group 2.
Among cisgender men, benign CL exhibited statistically significant differences from NVD in mean HNR (p = 0.004) and HNR SD (p = 0.002), and in mean HNR SD from laryngeal cancer (p = 0.027). Because of smaller sample sizes, there were no statistically significant differences in any feature among cisgender women.
The results indicate that HNR, especially HNR variability (HNR SD), may be useful for early identification of vocal fold lesions and could help distinguish between benign and malignant cases. Prior studies have established a link between acoustic parameters and laryngeal pathology, but this study provides additional evidence using a large ethically sourced data set.
Nonetheless, differentiating vocal fold lesions from other anatomical vocal pathologies like spasmodic dysphonia or uvular vocal fold paralysis (UVFP) remains challenging. The researchers point out how AI-based voice screening tools could allow for scalable, non-invasive triage of at-risk patients if they are developed with a larger and more diverse dataset. AI could facilitate earlier specialist referral and decrease the time from diagnosis to treatment, potentially even in telehealth or primary care settings.
Limitations of the study include the relatively small number of participants in some of the different sub-groups and incomplete lesion histories. Future research will focus on enrolling more participants, especially females, and incorporating additional acoustic features to improve diagnostic accuracy. The study highlights the growing potential of “audiomics”, the extraction and analysis of health information from voice to complement existing diagnostic methods in otolaryngology.
References: Jenkins P, Harrison R, Bedrick S, Karstens L, Hersh W. Voice as a biomarker: exploratory analysis for benign and malignant vocal fold lesions. Front Digit Health. 2025;7:1609811. doi:10.3389/fdgth.2025.1609811
