Hands are already busy
Holding the part, holding the fixture, holding someone else back. A voice is the one input that is always free.
Voice stop-request module · VSR-1
Say “stop.”
It stops.
RobotStop is safety hardware for physical AI. It bolts onto a robot you already own with one goal: a stop request in every spoken language and dialect on Earth. Two independent channels, all detection on the device, no wake word, no cloud, no network. Language coverage is expanding; each variety must be validated before it is listed as supported.
An extra layer, not a replacement. RobotStop does not stand in for your emergency stops, guarding, procedures or risk assessment — what that means.
- Stop latency
- ≤ 300 ms99th percentile
- Detection channels
- 2independent, cross-monitored
- Language ambition
- Allspoken languages + dialects
- Cloud services
- 0all detection on-device
01 Why it exists
In the second that matters, nobody finds the button.
An emergency stop only works if a hand reaches one. The person who sees the pinch point first is usually the person who can’t get there — hands full, wrong side of the cell, already on the floor. What everyone does instead is shout. RobotStop makes shouting work.
The button is somewhere else
Mushroom heads live on the cabinet and the pendant. Trouble happens at arm’s length, facing the other way.
Shouting is the reflex
People yell before they reach. Give that reflex a wire into the safety chain and you buy back the time it costs.
02 On a real arm
One word, mid-motion.
A 7-axis arm running a live task. No pendant, no keyboard, no operator standing at a cabinet. The word is spoken and the motion ends.
- Detected on the module — not on a PC, not in a data center
- Latched: it stays stopped when the talking stops
- Wires into the protective-stop input the robot already has
03 How it works
Two channels that don’t trust each other.
Redundancy only counts when the two halves can fail apart. RobotStop runs two complete detection paths that share no microphone, no power rail and no state — and either one of them can stop your machine. Both run the same detector, on purpose: one model, qualified once, held to one published bar.
-
01
Listen
Each channel has its own microphones and its own safety-manual microcontroller. No shared parts, and no general-purpose operating system anywhere in the stop path.
-
02
Decide
Both channels run the same detector — the one we qualify, at one operating point, against hundreds of hours of hard negatives. No wake word, no enrollment; calm or shouted, both count.
-
03
Stop
Either trigger de-energizes the outputs. The stop latches — silence does not clear it. Only a monitored rising-then-falling reset edge releases it.
-
04
Prove it’s alive
An acoustic self-test fires a test signal through the real microphone path on a schedule, and the two channels cross-check each other’s liveness. A dead mic becomes a declared fault in seconds, not a surprise on demand.
04 Languages
Every language. Every dialect.
A stop request in every voice.
We are expanding our ambition to every spoken language and dialect on Earth, including regional varieties and communities with little recorded speech. Local words, accents and pronunciations all belong in that work. Each variety needs its own recordings, evaluation and published evidence.
- Language entries
- 8,238Glottolog 5.3 catalog
- Dialect entries
- 13,706documented in the same catalog
- Combined entries
- 21,944families and bookkeeping excluded
- Our ambition
- Allspoken languages and dialects
How many is “all”?
There is no definitive worldwide total for languages plus dialects. Glottolog 5.3 identifies 7,674 spoken first-language languages; its broader catalog also includes signed, historical and other categories. Our 21,944-entry reference inventory counts language and dialect records together. It is neither a census of living spoken varieties nor a count of languages RobotStop currently detects. Dialect documentation is incomplete and classifications change.
Coverage today: the development word lists cover English, Mandarin Chinese and Korean. Word lists alone do not establish qualification. The worldwide scope is an expansion program; support must be established for the model and variety you deploy.
Signed languages are part of the world’s linguistic diversity and appear in the catalog. A microphone cannot recognize signing; signed-language support would require a separate sensing approach.
Drag to spin · search for a country
Official languages around the world
197 countries · official languages by country
United States
Official languages
- English
Regional designations in the reference
One marker per country. Official, de facto and working languages have separate labels. Regional designations appear separately where recorded.
Country language reference adapted from Wikipedia contributors under CC BY-SA 4.0, with linked corrections; locations from Natural Earth (public domain). Country data, scope and sources · Updated October 6, 2026. Language status is a reference, not a detector support claim.
Is your language included?
Check a language or dialect by its name, ISO 639-3 code or Glottocode. Catalog inclusion does not confirm detector support.
Sources and catalog download
Source: Hammarström, Forkel, Haspelmath & Bank, Glottolog 5.3 (2026), CC BY 4.0. Filtered and reformatted; retrieved October 6, 2026. Download all 21,944 entries (CSV) · Count method and attribution.
Can’t find your variety? Tell us its name and region. Catalog inclusion does not limit the expansion program.
Qualification follows each variety
Each release must name the languages, dialects and keywords it covers, with results for normal, raised and shouted speech. A parent language’s results do not qualify every dialect. Our protocol calls for at least 20 speakers and 1,000 positive utterances per acoustic condition for each variety claimed, plus confusable speech and noise testing with the full enabled language set.
Our goal: every voice heard. Every community included.
05 Versus building it yourself
You wouldn’t build your own e-stop. Why would this be different?
Nobody machines their own mushroom button, writes their own safety relay, or signs off on their own light curtain. They buy the part that arrives with a test report behind it. A voice stop is the same category of thing in newer clothes: the keyword spotter is a solved undergraduate project, the stop you are willing to put your name on is not, and the gap between those two is the entire product. The demo takes a weekend. The safety case takes a year.
| In-house build | RobotStop | |
|---|---|---|
| Redundancy | One mic, one board, one model | Two independent microphone channels, cross-monitored — either one alone stops it |
| Failure mode | Fails silent — you find out on demand | De-energize to stop; any detected fault drives outputs off |
| Reset logic | A boolean somebody clears in software | Latched stop, monitored reset edge, stuck-high rejected |
| Self-test | None. A muted mic looks identical to a quiet room | Periodic end-to-end acoustic test through the real mic path |
| False triggers | “Don’t stop,” “bus stop,” and the robot’s own voice all fire it | Tuned against hundreds of hours of hard negatives, incl. robot speech |
| Evidence | A video of the time it worked | A published test protocol and a witnessed report per release |
| Sign-off | You, explaining a side project to your safety officer | Documentation written for the person who has to sign |
| If it misses | That is on you | That is on us — it is the only thing we build |
Nobody wants to own the safety feature. Own the robot; let us own the stop.
The honest math: what building it in house costs, and when it is worth it →
06 What you can buy
Three ways to put it on your robot.
Start with one detector on one machine. Scale to a fleet when it has earned it.
Out of the box
RobotStop Detector
$10/ detector · your first one is free
About $1 of that is the hardware. Over 10,000 units it drops to about $1.10 a unit: the hardware plus $0.10 for us. See the math.
- Plug in, attach to the robot you already own
- Listening from boot — no setup, no account, no network, no wake word
- 3.3 V logic output, de-energize to stop — drive an optocoupler or a relay coil. The OSSD pair and NC dry relay ship with the certified module.
- USB-C: status, JSON per detection, and it types results as a USB keyboard with no driver
- Red / green verdict LED
- English, uncertified, one detection path. An evaluation unit, not a safety device
Custom services
Integration & Engineering
Quotedper robot
Scoped from your cell, your controller, your risk assessment.
- Full battery cutoff — the stop opens the pack contactor and removes power at the source
- Soft stop and hard stop — selectable outputs for controlled deceleration or immediate power removal
- Wiring, commissioning and sign-off support against your safety PLC
- ROS 2 driver integration, language and dialect evaluation, custom keywords
- Microphone placement and acoustic coverage survey for the cell
- Enclosure and mounting adapted to your machine
Fleet
Volume & Support
~$1.10/ detector over 10,000 units
Our ~$1 hardware cost plus $0.10 a unit. That’s all we make.
- Volume pricing over 10,000 units: about $1.10 each, on every unit in the order
- Staged rollout across a fleet, with spares held for you
- Qualification and compliance document pack for your files
- Priority support, response-time commitments, RMA handling
- Operator and maintenance training on the stop behavior
07 Straight about standards
What it is, and what it is not.
Safety products earn trust by being precise about their own limits. Here are ours, in the same words we put in the manual.
What it is
- A supplementary protective measure: one more stop-request input to the safety-related control system your machine already has.
- Designed to an ISO 13849-1 Category 3 architecture, targeting PL d, so certification is possible without a redesign.
- Frozen at release. No field learning, no drifting thresholds. Signed firmware and secure boot ship with the certified module.
- Entirely local. No network connection, nothing outbound, no continuous audio stored. The radio on today’s module is unused and disabled in firmware; the certified part will not have one.
What it is not
- Not an emergency stop device in the sense of ISO 13850, and never labeled or sold as one.
- Not a substitute for guarding, e-stops, interlocks, procedures, training or your risk assessment. It is added to them, never instead of them.
- Not yet certified as a safety component. Units shipping today are evaluation kits, and they go out in writing as such.
- Not a reason to stand anywhere you would not have stood before.
Qualification protocol v1
Our release-blocking thresholds and the method behind them are published, so you can hold us to them — or measure any other voice stop against the same bar.
- Detection, person near the robot
- ≥ 99.5%≤3 m, ≥+10 dB SNR
- Detection, busy floor
- ≥ 97%≤5 m, 0 dB SNR
- Detection, shout across the cell
- ≥ 90%at 10 m
- Nuisance triggers
- ≤ 1per 500 h of floor noise
- Stop latency
- ≤ 300 ms99th percentile
- Each channel alone
- Meetsthe same detection bar
Thresholds are met on the 95% lower confidence bound, not the point estimate. Every threshold is verified per firmware release and the report is witnessed by an independent party.
Safety notice
Keep every stop you already have.
RobotStop is a supplementary protective measure and is not an emergency stop device under ISO 13850. It is not certified as a safety component. Nothing about installing it justifies removing, bypassing or deferring an emergency stop, a guard, an interlock, a procedure or a training requirement — it is worth having only while all of those stay in place. Voice detection can miss a word; a cable, a host or a power supply can fail outright.
Get started
Put a stop on it this month.
Tell us the robot, the cell, and who signs off on safety. We will tell you exactly what it takes — including when the answer is that you don’t need us yet.