AI Can Read Subtle Cancer Signals in the Voice: Early Screening for Laryngeal Disease

Background: why voice matters for laryngeal cancer

Laryngeal (voice-box) cancer affects hundreds of thousands worldwide. In 2021 roughly 1.1 million cases of cancers of the larynx and nearby structures were recorded globally, with nearly 100,000 deaths. Standard diagnosis relies on specialist procedures such as video nasal endoscopy and tissue biopsy—effective but invasive and dependent on access to trained clinicians.

Study design and methods

Researchers at Oregon Health & Science University and Portland State University analyzed 12,523 voice recordings from 306 North American participants to explore whether vocal fold lesions—both benign and malignant—leave measurable signatures in the sound of the voice. Using machine learning models to extract acoustic features, the team searched for patterns that distinguish cancerous lesions from benign growths and other voice disorders.

Key findings: harmonic-to-noise ratio as a marker

The study identified specific acoustic differences in male voices. The harmonic-to-noise ratio (HNR)—a metric that quantifies the balance between periodic tone and aperiodic noise in the voice signal—was particularly informative. Models trained on HNR and related features separated men with cancerous vocal fold lesions from those with benign lesions or other voice disorders, at levels undetectable to the unaided ear.

"To move from this study to an AI tool that recognizes vocal fold lesions, we would train models using an even larger dataset of voice recordings, labeled by professionals. We then need to test the system to make sure it works equally well for women and men," says clinical informatician Phillip Jenkins of Oregon Health & Science University.

Limitations: sex differences and dataset size

The team did not find statistically significant acoustic markers in women within this dataset. That does not rule out detectable features in female voices; it indicates that larger, more diverse datasets and additional model tuning are required before a universal screening tool can be implemented.

Related technologies and future prospects

Voice-based diagnostic tools—driven by speech processing, digital signal analysis, and machine learning—are already being piloted for several conditions, from COVID-19 screening to neurological disease monitoring. A validated voice-screening tool for vocal fold lesions could enable non-specialist clinicians, telemedicine platforms, and population-level screening programs to triage patients more quickly and refer high-risk individuals for definitive ENT evaluation.

"Voice-based health tools are already being piloted. Building on our findings, I estimate that with larger datasets and clinical validation, similar tools to detect vocal fold lesions might enter pilot testing in the next couple of years," Jenkins added.

Clinical and ethical considerations

Before deployment, prospective tools must undergo clinical validation across sexes, ages, languages, and recording conditions to avoid biased performance. Regulatory approval, privacy safeguards for voice data, and clear referral pathways for positive screens are essential to convert acoustic biomarkers into safe, equitable public-health tools.

Conclusion

Machine learning analysis of voice recordings has revealed subtle acoustic markers—notably the harmonic-to-noise ratio—that can distinguish cancerous vocal fold lesions in men. While findings are preliminary and female-specific markers remain elusive in this dataset, the research demonstrates a promising path toward non-invasive, scalable screening for laryngeal cancer. With larger labeled datasets, clinical validation, and careful attention to equity and privacy, voice-based screening tools could accelerate diagnosis and improve outcomes for patients with voice-box disease.