Kintsugi's voice test for depression ran out of runway waiting on FDA clearance

Voice-biomarker API for health systems and telehealth: scores PHQ-9/GAD-7 depression and anxiety risk from a short speech sample, content-blind and language-agnostic.

That vocal biomarkers — pitch variation, pacing, pauses — can flag depression as reliably as a PHQ-9 questionnaire, and that regulators would clear it as a medical device

$8M seed in 2021 and a $20M Series A in early 2022, about $30M total.

Grace Chang, who grew up in Taiwan, and Rima Seiilova-Olson, who grew up in Kazakhstan, met at a 2019 hackathon. Both engineers had been through depression and both had hit the same wall getting care. They framed the problem as triage: with roughly 280 million people depressed worldwide and too few clinicians, providers had no fast way to see who was struggling.

The pair raised $8M in 2021 and a $20M Series A in early 2022. The model, trained on tens of thousands of consented voices, scores the PHQ-9 (0–27) and GAD-7 (0–21) scales from speech patterns alone — it deliberately ignores what is said, which made it language-agnostic. Kintsugi ran a clinical study with the University of Arkansas for Medical Sciences, signed contracts with large US healthcare companies, and pursued FDA De Novo classification while selling as clinical decision support under the 21st Century Cures Act.

FDA clearance never arrived. In February 2026 Kintsugi closed, and rather than shop the models to buyers it published its technology and research as open source (Behavioral Health Business, 2026-02-11). Commentators noted the voice models could get a second life in deepfake-audio detection.

Selling as 'clinical decision support' under the 21st Century Cures Act was a gray zone that narrowed: health systems and payers ultimately want a regulated device, not a screening heuristic.

FDA validation for voice AI means controlled, demographically broad datasets — a spend that compounds every quarter while the clearance, and the revenue behind it, waits.

Voice biomarkers shift with accent, age, comorbidities and even microphone quality, so the false-positive risk was exactly the failure mode regulators are least willing to tolerate.

Seven years is longer than a venture runway: the bet needed the FDA to move on startup time, and De Novo review does not.

In regulated health AI the model is the easy part; multi-demographic clinical validation is the cost center. Price the regulator into the runway, or ship where software is allowed to stay software.

As of April 2026 the technology and research are public: Behavioral Health Business reported the open-source release on 2026-02-11, so labs and rivals can build on the dataset work instead of starting from zero. Startup Fortune noted the voice models could be repurposed to detect deepfaked audio, since the patterns that betray mood also betray synthetic speech. The founder's argument — that healthcare AI struggles to reach regulated revenue before venture money runs out — became the story's real echo in the trade press.