Built on Bee. Not a diagnosis.
Speech is a vital sign. Start measuring yours.
Bellwether reads the transcripts your wearable already makes, keeps nine language numbers and throws every word away, learns how you normally speak, and tells you when that changes against your own past.
The band is this person's own ordinary range, learned from their own earlier days. Nothing here is compared with anyone else. That is not a preference: measured on 349 real recorded conversations, the gap between two different people is smaller than one person's own day-to-day range on all nine measures, so comparing you with a population would be asking a detector to resolve a difference smaller than the noise.
Bee, Amazon Bedrock, AWS, and the Model Context Protocol. Open source under MIT.
Your watch knows when your heart skips. Nothing knows when your brain does.
Speech is the brain's only continuous, observable output. When something changes, speech is usually where it shows first: more pauses, simpler sentences, lost words. Doctors know this. They also hear a patient for a few minutes a year, so the change hides in plain sound, for years.
3.5 years
from first symptoms to a dementia diagnosis, on average
Early stage only
where the newest Alzheimer's treatments are approved to work
40 to 80%
of what a doctor says is forgotten immediately
85 to 89.6%
accuracy of speech-based detection on the field's public benchmark
How it works
Four steps. The words never leave your machine.
Step 1
Wear
The Bee wristband transcribes your day. You already have it on.
Step 2
Measure
On your own computer, each day of speech becomes nine numbers. The words are discarded.
Step 3
Learn
A week of your normal. Not anyone else's. Your accent, your pace, your habits are the baseline.
Step 4
Show
Stable, watch, or discuss, with the reasons and the dates, and a one-page report for the appointment.
A week with Bellwether
What the wearer actually does. Almost nothing, on purpose: a measurement you have to work at is one you will stop taking.
Day one
One command
bellwether-ingest mine reads your Bee, turns each day into nine numbers on your own computer, sends only those, and opens your page.
Every day
A two-minute check-in
Say "Bellwether check-in" and the date, then talk to the band on your own. Or let the assistant host it: it asks three standard questions through the speaker, the band records your answers, and the host's words are removed before anything is measured.
Day eight
The first comparison
Seven days taught it your normal. From here, each day is read against them, and the coaching note says what to do, not just what moved.
Any day
Ask out loud
How have I been sounding? Through Alexa+ or any assistant that speaks MCP, answered from the same nine tools.
Appointment day
Print one page
What changed, when, which measures moved, the notes you added, and in full: this is not a diagnosis.
A live baseline, on the real engine
This is the deployed service. The person is simulated; the extractor, the baseline engine and the tiers are the product.
The evidence
Every number on this page, with its source. Each was checked at the source during the build.
Time to diagnosis in dementia: systematic review with meta-analysis
Orgeta et al., Int J Geriatr Psychiatry, 2025
3.5 years on average from symptom onset to diagnosis across 13 studies and 30,257 participants; 4.1 years for young-onset.
Lecanemab and donanemab indications
U.S. FDA, 2023 and 2024
Both approved only for mild cognitive impairment or mild dementia due to Alzheimer's disease, with no data on starting later.
Patients' memory for medical information
Kessels, J R Soc Med, 2003
40 to 80 percent of medical information is forgotten immediately; almost half of what is remembered is incorrect.
The ADReSS challenge
Luz et al., Interspeech, 2020
Best systems separate Alzheimer's from control speech at 85 to 89.6 percent on a benchmark balanced for age and gender.
Linguistic ability in early life and Alzheimer's disease in late life
Snowdon et al., JAMA, 1996
Low idea density in writing at age 22 predicted Alzheimer's disease at autopsy some 58 years later. The Nun Study.
Linguistic features identify Alzheimer's disease in narrative speech
Fraser, Meltzer and Rudzicz, J Alzheimers Dis, 2016
The linguistic profile this product measures: reduced diversity and syntactic complexity, more pronouns and filled pauses.
We publish our own validation
docs/EVIDENCE.md, section 6
Two synthetic personas through the production extractor and engine: 47 quiet days with no false alarms, and a gradual change caught two days in. Plus what it does not show, stated plainly.
Who it is for, and what it would change
Three people meet the same three and a half years from different sides. None of them has had a record of the one thing that was changing all along.
Someone who has started to wonder
Sixty or older, wearing a Bee anyway. Wants a record, not a verdict. Gets a baseline that is only theirs, compared with nobody, and one page when they are ready to ask.
The family member who notices first
Lives far away and hears it on the phone, with no way to say it. Gets a shared record of change with dates on it, instead of an argument about memory.
The clinician with twelve minutes
Hears a patient for a few minutes a year. Gets nine measures across months, which way each moved, and the ordinary explanations already listed.
What people say they want, in their own answers
79%
of Americans 45 and over would want to know they had Alzheimer's before symptoms, or before symptoms interfere with daily life.
Alzheimer's Association survey of 1,702 adults, 2025
66% vs 14%
want brain health guidance from their doctor first, yet only 14% have ever discussed brain health with their physician.
Alzheimer's Association survey of 3,829 adults, 2026
People want to know early, they want to hear it from their doctor, and the conversation almost never happens. The one-page report exists to start it.
The case, in three verified steps
Step 1
92%
of expected mild cognitive impairment cases in US Medicare were never diagnosed, 2015 to 2019. Detection was about half as high for Black and Hispanic beneficiaries.
Bellwether uses no population norm: each person is compared with their own past.
Mattke et al., Alzheimer's Research & Therapy, 2023
Step 2
7.6 years
before diagnosis, language from people who were still cognitively normal already predicted Alzheimer's onset (AUC 0.74, Framingham Heart Study).
Written picture descriptions, compared across people. It shows the timing, not Bellwether's accuracy.
Eyigoz et al., EClinicalMedicine, 2020
Step 3
Every year
Medicare's Annual Wellness Visit already requires an assessment to detect cognitive impairment, and names no single tool.
The doctor page has a section written for that visit: observation over time, never a screening result.
Cordell et al., Alzheimer's & Dementia, 2013
The economics, in four numbers
17%
of net benefit lost for every year a diagnosis is delayed
Barnett et al., BMC Neurology, 2014
$405,262
lifetime cost per person with dementia, about 70% borne by families
Alzheimer's Association, 2026
$49.99
once, for a wristband people are buying anyway
Bee
Pennies
per person per month to run, measured on the live deployment
AWS Cost Explorer
What is measured, and by whom
The delay costs about 17 percent a year: measured, by others. Language features separate Alzheimer's from control speech at 85 to 89.6 percent on the field's benchmark: measured, by others. Those features come out of ordinary wearable transcripts at pennies, daily, with no clinic: measured, by us, on the production pipeline.
What connects them into a shorter delay is a trial nobody has run. We would rather say so than invent a number. The study to run is written, with its endpoints fixed before any data.
Beyond this hackathon
A family at home: one command, a two-minute check-in a day, a page for the appointment.
Someone who will not use a screen: the assistant hosts the check-in by voice, asks three questions, and reads back where they stand. No app to open, nothing to read.
The yearly visit: the doctor page has a section written for Medicare's Annual Wellness Visit, which already asks whether anything has changed.
A memory clinic between annual visits: the whole stack deploys from one CDK command into the clinic's own account, and only nine numbers a day ever reach it.
Any transcript, not only Bee: the open-source speech-vitals package reads any speaker-labelled transcript, so the engine is not tied to one device. English only today, and the FAQ says so.
What we store, and what we never do
Stored
- Nine numbers per day, such as vocabulary variety and idea density
- How many words were spoken that day
- Notes you add, such as travel or poor sleep
- Spoken check scores
Never stored
- Transcripts, sentences, or any words
- Audio of any kind
- Other people's speech
- Anything that could reconstruct a conversation
Check it yourself
This sends one day to the real intake with an ordinary morning's transcript attached, the kind a wearable hears, full of a name, a doctor, a family member, an address, an account number and a medication. Then it reads back what was kept and searches it for every one of them.
Bellwether is a general wellness tool. It does not diagnose, treat, cure, or prevent any disease, and it is not a medical device. Changes in speech have many everyday causes, including sleep, stress, medication, and mood. Share the report with a clinician you trust.
Use it with your own Bee
The person above is simulated so you can see a year at once. This is how you would run the same thing on your own speech.
Connecting your own Bee
Five steps. The first four happen on your own computer, which is where your words are and where they stay. Only the last one uses the network, and it carries nine numbers a day.
Or the whole thing, in one command
Steps 1 and 2 first, then thisReads your check-ins and the band's conversations, turns each day into nine numbers here, sends only those, claims your profile with a key that stays on your computer, and prints the link to your own page. Nobody else can post a day as you after that. If Bee is not logged in it says so in the words the Bee app uses. Rehearsed in tests for every way it can stop, and for the one thing it must never do: print a word you said.
bellwether-ingest mine --profile <your-name>No Bee? Use the meetings you already have
Your computer, words stay hereZoom and Teams transcripts (.vtt), Otter exports (.txt), or any “Name: text” file. Run it once without a name to see who spoke, then again with yours. Only your own turns are kept, and only nine numbers a day leave your computer.
bellwether-ingest file meetings/*.vtt
bellwether-ingest file meetings/*.vtt --speaker "Your Name" --post <your-name>What that command does, step by step
1Let your computer read your Bee
Your BeeInstall Bee's own command line tool, then in the Bee app turn on Developer Mode by selecting the version number five times in Settings.
npm install -g @beeai/cli bee login2Install Bellwether's reader
Your computer, words stay hereIt runs on your machine, not ours. The feature extractor is the open source speech-vitals package from PyPI.
git clone https://github.com/usv240/bellwether cd bellwether pip install -e "packages/speech-vitals[nlp]" -e apps/ingest python -m spacy download en_core_web_sm3Rehearse the first run
Your computer, words stay hereRead only and safe to repeat. It checks the tool, the login, whether anything has been recorded, whether a real day reduces to nine numbers, and whether that day has enough speech to judge. It stops at the first thing that is actually wrong and says what.
bellwether-ingest firstrun4Turn your days into numbers
Your computer, words stay hereEach day of your speech becomes nine numbers. The words are read, measured and discarded here. Bee does not yet say which speaker is you, so there are two streams. A check-in is you: two minutes talking to the band alone, which you declare yours by opening out loud (Bellwether check-in, October 17) or with a short voice note the session opens with. The first command reads only those. The second is the band's everyday conversations with sessions over an hour and anything Bee calls media left out, which is a filter rather than your voice alone.
bellwether-ingest pull --checkins --out voice.json bellwether-ingest pull --backfill --max-minutes 60 --out days.json5Send the numbers
Leaves your computer, numbers onlyThis is the only step that leaves your computer, and by now there is nothing in it that could reconstruct a sentence. Open the file first if you want to see for yourself.
curl -X POST https://bppni6dpuntpbynfydk52gexue0xulzh.lambda-url.us-east-1.on.aws/v1/profiles/me-voice/days \ -H "content-type: application/json" -d @voice.json curl -X POST https://bppni6dpuntpbynfydk52gexue0xulzh.lambda-url.us-east-1.on.aws/v1/profiles/me/days \ -H "content-type: application/json" -d @days.json
Where this actually stands: every step above has run on the builder's own band since 2026-09-29, and bellwether-ingest evidence writes the public record of it, counts and the nine numbers and not one word of speech. What Bee does not yet give is who was speaking: every ambient utterance is labelled Unknown, so the wearer's own speech comes from check-ins they start on purpose, and the conversation stream is shown as the room's, not the wearer's. Both are on the dashboard, labelled.
Ask Alexa+ how you have been sounding
Nobody opens a dashboard to ask this. They ask it out loud, in the kitchen, after wondering for a while. So the spoken surface is a real one: a self-hosted MCP server an assistant can hold a session with. Press it and watch.
Asking out loud, instead of opening a dashboard
The question this product answers is one people ask out loud: how have I been sounding lately. Bellwether answers it through the Model Context Protocol, which is how Alexa+ talks to tools. Press this and your browser will hold a complete session with the deployed server, the way an assistant would. Every row below is the server's own answer.
For developers
The API accepts feature rows, not text. Run the open-source speech-vitals package where the transcript is, and post the numbers. That is a stance, not a limit.
pip install "speech-vitals[nlp]"
python -m spacy download en_core_web_sm
speech-vitals analyze transcript.jsonl --speaker speaker_1 --out days.json
curl -X POST https://bppni6dpuntpbynfydk52gexue0xulzh.lambda-url.us-east-1.on.aws/v1/profiles/me/days \
-H "content-type: application/json" \
-d @days.jsonEndpoints. GET /v1/profiles/{id}/summary, /assessments, /report, /weekly-note; POST /days, /annotations, /checks. Interactive docs at /v1/docs.
MCP. https://bppni6dpuntpbynfydk52gexue0xulzh.lambda-url.us-east-1.on.aws/mcp, spec 2025-11-25 over Streamable HTTP, nine tools, two of which host the check-in itself.
Resilience. /api/resilience reports every degradation path.
Bee. Ingestion calls bee conversations, bee now, bee changed, bee stream --json and bee sync in code.
About the build
Tracks
Bee (Wearable AI), with the MCP server as a genuine Alexa+ surface. Mini challenges: AWS Builder and Open Source.
AWS, and why
Bedrock phrases the weekly note through a three-model ladder that falls to a deterministic template. DynamoDB holds feature rows only. Lambda runs the API and MCP server without spaCy. CDK, S3 and CloudFront for the rest.
Open source
speech-vitals (MIT): nine language features per day with their literature basis, features only, never words. The engine, the Bee adapter and the service are MIT too.
The questions a careful person asks
Is this a diagnosis?
No. Bellwether describes change over time in one person against their own baseline and lists its reasons. It does not diagnose, screen for, treat or prevent any condition, and it is not a medical device. It gives you a record to share with a clinician you trust.
Who can see my data?
You. Nothing that could reconstruct a sentence is stored anywhere: transcripts are turned into nine numbers a day on your own machine and discarded. The service receives only the numbers, and a test proves that any field carrying text is dropped at the door.
What if I have an accent, or speak two languages?
Your baseline is you. Bellwether never compares you with a population, so an accent or a quiet manner is simply what normal looks like for you. Cross-language support is on the roadmap; the feature extractor is English-only today.
Why should a healthy thirty-year-old care?
Medication fog, poor sleep, illness and recovery all show in speech, and a long baseline is the most valuable thing this can give you. The earlier the baseline starts, the more a change later in life has to be compared with.
What does Bee record?
Bee transcribes conversations near the wearer, and today it labels every speaker Unknown, so it cannot say which words were yours. Bellwether therefore keeps two streams and labels them: your check-ins, which you declare yours out loud and record alone, and the band's conversations, which are the room. No audio reaches Bellwether at all.
Can it detect a stroke?
No. It is not an emergency product and reacts on the scale of days, not minutes. If speech changes suddenly, call emergency services.
Why does it say the person is simulated?
Because they are. The live baseline on this page is generated speech for one invented person, run through the real extractor and the real engine, so you can see eight weeks including a change from week six. Everything after the words is the product.
Can an assistant ask it for me?
Yes. Bellwether is a Model Context Protocol server, so a voice assistant or AI agent can ask how you have been sounding, read the report, run the spoken check with you by voice, or host tonight's check-in by asking the three questions itself. The assistant gets the same nine numbers-only tools and nothing else.