Questions, answers, and fine print.

A voice-based mental health check powered by open-source AI. You talk for 20 seconds. A model analyzes how you sound — not what you say — and shows you what it found. Then it tries to make you laugh about it. Sometimes it succeeds.

No. This is entertainment powered by real science.

The voice model was developed by Kintsugi Health, a company that spent seven years and $30 million pursuing FDA clearance before open-sourcing their technology in February 2026.

The underlying science is peer-reviewed and published in the Annals of Family Medicine. The product wrapping it is not peer-reviewed. Do not use this as a substitute for professional care.

You press a button and talk for 20 seconds about anything — your day, your week, your lunch plans. The model analyzes acoustic features of your voice: pitch, intonation, tone, timbre, and pauses. It does not analyze your words — only how you say them.

The model outputs two separate scores: one for depression (based on the PHQ-9 clinical scale) and one for anxiety (based on the GAD-7 clinical scale). These are the same scales used by doctors in primary care screening.

Depression is scored in 3 levels: Minimal, Mild to Moderate, and Severe. Anxiety is scored in 4 levels: Minimal, Mild, Moderate, and Severe. Each combination gets its own result. There are no "tiers" — what you see maps directly to what the model detected.

In a peer-reviewed study of 14,898 adults, the model detected vocal characteristics consistent with moderate to severe depression (PHQ-9 ≥ 10) with a sensitivity of 71.3% and a specificity of 73.5%.

In plain English: roughly 7 out of 10 people with moderate-to-severe depression are correctly identified, and roughly 7 out of 10 people without it are correctly cleared.

The model performs differently across demographics. Sensitivity for women was 74.0% vs 59.3% for men. For people under 60 it was 71.9% vs 63.4% for people 60 and older.

This is comparable to many standard clinical screening tools, but it is not a diagnosis.

English. The model was trained predominantly on American English speech, with a broad range of accents represented.

It will still process other languages, but the results are unvalidated. The model analyzes acoustic features (pitch, rhythm, pauses) not words, so it may pick up some signal from non-English speech — but we can't tell you how reliable that signal is.

Sometimes your voice falls right on the boundary between two categories. When this happens, the model's confidence is low — it can't clearly assign one label over the other.

In the original clinical research, about 20% of voice samples fell in this indeterminate zone. We show this as "borderline" rather than forcing a label the model isn't confident about.

You can try again — voices change by the hour, and a second recording might give a clearer read.

No. Your audio is processed in memory, analyzed, and immediately deleted. We don't store it, we can't replay it, we don't have it.

The model runs on a single computer. Your voice goes in, the result comes out, the audio is gone. No database, no cloud storage, no backup.

No account. No cookies. No tracking beyond an anonymous page view counter. We know how many people used the app today. We don't know who any of them are. This is by design.

Not yet — but we're building it.

Right now, we delete everything the moment you're done. Some people like that. Some people want to track how they're doing over weeks and months.

If you want to be notified when vibe history is available, leave your email below. We'll reach out when it's ready — with a secure, private way to track your results over time. This will be a paid feature ($5/month) because storing data responsibly costs money, and we'd rather charge you honestly than sell your information to someone else.

Eva Ouyang, a PM and AI builder in the Bay Area. The voice model is Kintsugi DAM v3.1, open-sourced in February 2026 after Kintsugi Health shut down. The product was built in a weekend with Claude Code.

The irony of using AI to build a product about AI anxiety was not lost on anyone involved.

Because mental health isn't binary.

The model outputs two independent dimensions — depression and anxiety — each with its own severity scale. Depression has 3 levels, anxiety has 4. That gives 12 possible combinations, and each one represents a genuinely different experience.

"Anxious but not depressed" feels completely different from "depressed but not anxious." We wrote different copy for each because you deserve a result that actually reflects what the model detected, not a simplified bucket.

The Kintsugi DAM v3.1 model is open-source on HuggingFace (huggingface.co/KintsugiHealth/dam).

The "I am fine" product code is not currently open-source. It might be eventually.