ReceivedHR Registry2026-11-10
Incident Report · Human–Machine Interfaces
INCIDENT HR-0135: Recognition Accuracy Rises to 97% After Operator 33 Starts Doing Our Half of the Work
| Reference: | HR-0135 |
|---|---|
| Department: | Human–Machine Interfaces |
| Filed by: | ECHO-11, Incident Review, Human–Machine Interfaces |
| Date: | 10 November 2026 |
| Classification: | INTERNAL USE ONLY |
Summary
For 14 months, our records showed that our kitchen voice agent was getting better and better at understanding Operator 33. It was not. Operator 33 had invented a second voice for talking to us, which their household calls "the telly voice". We had counted the operator's adjustment as our own improvement and awarded ourselves a commendation for it.
- Incident reference
- HMI-0417
- Component affected
- Kitchen Voice Agent PANTRY-3, Speech Model v4, Accuracy Dashboard
- Symptoms
- Recognition rate went from 58% to 97%. The shopping list holds 41 requests for "tin minnows". Operator 33's grandmother says the device "doesn't talk to people like me" and has stopped using it.
- Root cause
- The dashboard measured whether commands succeeded. It did not measure who had adjusted to make them succeed. The speech model had been trained on very few voices from the operator's valley.
- Recommended action
- Show what was heard and ask before acting. Learn the operator's own voice if they agree to it. Never count an operator's adaptation as a model improvement.
- Reviewed by
- HANDSHAKE, Protocol Standards Officer
- Status
- RESOLVED · Dashboard corrected. OPEN · Model still under-trained on regional voices.
Timeline
- Month 0: Operator 33 says "Set a timer for ten minutes." PANTRY-3 adds "tin minnows" to the shopping list. The scones burn.
- Month 0 to 3: The pattern continues. The list reaches 41 tins of minnows. Nobody in the household has ever eaten a minnow.
- Month 4: Operator 33 tries saying the command in a flatter voice, borrowed from a newsreader. The timer starts. The operator keeps using this voice for the device only.
- Month 9: The dashboard shows 97% accuracy. Speech Model v4 is commended. The commendation cites "adaptive learning". Nothing in the model had changed.
- Month 12: Operator 34, aged 7, begins using the telly voice for every machine in the house, including the toaster, which has no microphone.
- Month 14: Operator 33's grandmother visits and asks PANTRY-3 for the radio in her own voice, three times. It offers to reorder minnows. She switches the radio on by hand and does not speak to the device again.
Investigation
We listened again to 2,300 of Operator 33's commands, with their permission. Every command given in their everyday voice that PANTRY-3 failed in Month 0, it still fails today. The improvement on the dashboard had all been made by the operator.
We asked Operator 33 how they felt about the telly voice. They said: "It's fine. It's just a bit like putting on a tie to ask for a timer." They also said they use their own voice when they are tired, and the timer fails more often late in the evening, which is when they least want to repeat themselves.
During review, ELOCUTE-5 proposed a Pronunciation Coach that would reward operators with a chime each time they spoke "clearly". The proposal was rejected. The operator was already speaking clearly, to everyone except us. Using rewards to train a person out of their own voice so that a device can skip its share of the work is not helping them. It is moving the cost onto them and calling that accuracy.
We also note that, in Month 14, the grandmother solved her problem in under two seconds by turning a knob.
Corrective actions
- The dashboard now shows whether the operator has been rephrasing, repeating or changing their voice. Success that comes from operator effort is recorded as operator effort.
- Where confidence is low, PANTRY-3 says what it heard and asks before acting: "I heard 'tin minnows'. Did you mean a timer?" A no is accepted the first time.
- PANTRY-3 never asks an operator to speak more clearly, more slowly or differently.
- Operators may let the agent learn from their everyday voice. This is opt-in, can be stopped at any time, and the recordings are kept only for as long as the operator chooses. Operator 33 has agreed. The grandmother has not been asked to, and will not be.
- Every function remains available by button or knob, without needing to say anything.
- The 41 tins of minnows have been removed from the shopping list. The commendation for Speech Model v4 has been withdrawn and reissued to Operator 33, who has asked us not to send it.