with guest Henry O’Connell, CEO and Co-Founder, Canary Speech
Season 2: Episode 2
Dr. Heather Bassett, Host of AI Amplified, welcomed Henry O’Connell, CEO & Co-Founder of Canary Speech to the show. Canary Speech traces back to an eight-hour lunch in a Provo, Utah bagel shop between two friends of forty years, both in their sixties, both between projects. Henry O’Connell asked Jeff Adams — the engineer who’d led development of the Amazon Echo and Dragon NaturallySpeaking — a question he thinks everyone should ask people they care about: what’s left on the road in front of you? Jeff’s answer was that he’d always wanted to apply speech technology to human health. Henry’s response was immediate: why don’t we do that?
The idea crystallized around something Henry had noticed in his own family. His daughter, walking across the room after soccer practice, told him everything about her day through gait alone, before she said a word. After she moved out, phone calls became his only window. And the same read held, based purely on voice, regardless of whether she answered the question every parent asks with the answer every parent gets: fine. The pattern had nothing to do with the words themselves. That observation became Canary Speech’s founding thesis: build something that validates what a trained ear already senses, and normalizes it across an entire population.
Getting there took longer than expected, partly because the field’s first 25 years of work rested on a flawed foundation. Canary Speech initially followed the standard speech-pathology approach: reading a structured paragraph aloud, seeded with challenge words. But reading isn’t the same cognitive act as speaking; it’s closer to acting. Modeled against diagnostic outcomes, structured speech produced accuracy in the high 50s and low 60s. Conversational, unstructured speech pushed that into the 80s — a gap wide enough that the team abandoned analyzing words entirely.
What replaced it measures how speech is physically produced rather than what it expresses: 2,590 distinct features captured in a 25-millisecond window sliding every 10 milliseconds, all tied directly to central-nervous-system function — vocal cord movement, rate of change, the underlying articulatory mechanics beneath any language. That works out to roughly 15.5 million data points per minute, against the 800 or so available from word-based analysis. Language is expressed differently across cultures, genders, and dialects, but it’s created by the same neural machinery in everyone. Measuring at that foundational level, rather than where language gets expressed, is what gives the technology a real shot at generalizing across diverse populations — a distinction Henry credits with the company’s more recent work, including a current study on multiple sclerosis identification.
Henry’s approach to earning clinical trust starts with a simple premise: build it with the people who will use it, not for them. Canary Speech runs 17 ongoing clinical trials and partnerships, including Harvard’s Beth Israel, Mass General, NYU, Hackensack Meridian, and roughly two dozen hospital centers across the Midwest, plus work spanning Japan, South America, Saudi Arabia, and Europe — every deployment under an IRB-approved protocol. For the company’s first nine years, its technology was research-only, never sold. That rigor has produced two published studies with Harvard, a study with the NIH, and a paper in The Lancet within the past year, credibility that lets Canary Speech partner with organizations like Microsoft, Zoom, and Teladoc rather than try to sell directly into health systems that already rely on those platforms. The company is also CPT-code eligible and insurance-reimbursable as a clinical decision support tool, audited regularly against FDA guidance.
That partnership model also solves healthcare AI’s usual adoption problem: Canary Speech runs as a toolbar component inside tools clinicians already use, like Dragon Copilot, with results surfacing directly in Epic’s Haiku app — no separate login, no second device, often no extra consent click. Dr. Bassett, drawing on her own clinical background, connected this to something physicians and nurses already do instinctively the moment they walk into a room: read a patient’s condition from cues that never make it into the chart, often within a 15-minute visit too short to properly screen for something a patient may be actively trying to hide, like early-stage dementia. The value isn’t replacing that clinical judgment; it’s making an instinct clinicians already trust into something measurable and repeatable at scale.
Mental health data carries a heavier protection burden than most clinical information, and Henry was direct about how Canary Speech handles it. Audio arriving from Dragon Copilot or Zoom comes in de-identified and encrypted, satisfying HIPAA and HITRUST requirements, and Canary Speech never stores the audio itself; it exists only on the healthcare partner’s side, under their own protections. That infrastructure has let the company expand into a second market: workplace wellness, where individuals can self-assess and get pointed toward resources while employers see only de-identified, population-level dashboards. Henry’s example was specific — a company with roughly a thousand locations noticing one site trending below average on behavioral health, and responding not with individual intervention, but with something as simple as a weekly summer barbecue, a way of signaling that the group is seen without singling anyone out.
Henry traces his own conviction about healthcare access back to a childhood shaped by poverty, and to a doctor and dentist who made house calls simply because that’s what the era’s version of care looked like. His belief now is that access to care shouldn’t depend on geography or the number of clinicians a system has the capacity to train, especially given how far training capacity lags behind an aging, growing population. Tools that act as force multipliers — giving clinicians objective information in real time, on whatever device is already capturing audio, from an Apple Watch to a Samsung Watch — are, in his view, the most realistic path to closing that gap. Dr. Bassett closed the conversation by echoing that same belief back: that extending limited clinical resources to the actual point of care, wherever that point happens to be, remains one of the more meaningful things AI in healthcare can still accomplish.
Healthcare AI is moving fast, and not all of it lives up to the hype. On AI Amplified, Dr. Heather Bassett, Chief Medical Officer at Xsolis, sits down with industry leaders and AI experts to separate what’s real from what’s noise. Each episode explores the innovations paying off, the challenges still ahead, and the lessons shaping what comes next, all with an eye toward restoring the joy in medicine.
AI Amplified is brought to you by Xsolis, an AI-driven healthcare technology company transforming the payer-provider dynamic and uniting these two crucial stakeholders like never before. Since 2013, Xsolis has been breaking down silos to eliminate friction, waste, and unnecessary manual work in healthcare.
Join us in solving some of the biggest challenges in healthcare. Request a consultation to learn more about Xsolis.