Canary Speech Explores Voice as a Health Signal
A podcast episode features Canary Speech CEO Henry O'Connell discussing how vocal traits may offer useful health information. He explains that the company found unstructured conversation more diagnostically accurate than scripted reading in its models. The discussion centers on measuring how speech is produced rather than analyzing the words themselves.
Canary Speech grew from a conversation between Henry O’Connell and Jeff Adams, a speech-technology engineer, about applying such tools to health. O’Connell’s observations of his daughter—first through movement, later through phone calls—suggested voice itself might carry signals independent of wording. Early models using read passages scored in the high 50s to low 60s; unscripted talk reached the 80s, prompting the company to stop analyzing words.
Its approach instead tracks physical speech production: 2,590 features in a 25-millisecond window updated every 10 milliseconds, about 15.5 million data points per minute versus 800 from text-focused methods. The firm spent nine years research-only and runs 17 ethics-board-approved trials and partnerships across hospitals and regions, with published work involving Harvard and the NIH.
If voice-based tools prove reliable, they could change how clinicians screen and monitor conditions, especially for people with limited access to specialists or those who find scripted tests difficult. Patients may benefit from faster, less intrusive assessments, while employers, insurers, and health systems could face new questions about consent, data protection, and interpreting probabilistic signals. The technology may also reshape remote care, but its value will depend on validation across diverse populations and careful communication of uncertainty.