Misrecognised voice commands cause unintended state changes

What it is

Speech recognition errors are not limited to "didn't understand". The worse class is "understood wrong": an utterance is recognised as a different, valid command and executed, changing device state without anyone intending it. The full fault chain has two steps: a false wake (the assistant starts listening though nobody addressed it) plus a misrecognition (whatever it hears resolves to the wrong command). TV dialogue, family conversation, and noises from the next room can supply both.

Unlike tapping the wrong button in an app, a misrecognised command often looks unprovoked to everyone else: nobody knew anyone was talking to the speaker, and the light goes off or the door locks — to bystanders, the device acted on its own, which is a far worse starting point for diagnosis than "pressed the wrong thing".

Why it happens

The structural reason voice is more dangerous than buttons: a button's physical placement bounds its blast radius — press the wrong switch and at worst that switch is wrong. A speech error lands in semantics: recognition can output any valid command, and the system has no grounds to doubt a sentence that is grammatically pristine. Recognition confidence drops under noise, accents, and far-field pickup, yet the system still emits the top candidate and executes — there is no built-in default of "when unsure, do nothing".

False wake expands the attack surface: a standby speaker is always listening for its wake word, with TV audio and conversation streaming through the front-end detector. Wake detection is a threshold decision — set it low and false wakes multiply; set it high and genuine calls go unanswered. Multiplied together, misexecution is not a tail event but a daily occurrence in noisy households.

Consequences are amplified by device class: voice gets used for high-impact devices — locks, sockets, appliances — so one misrecognition becomes a physical-world action; and where voice is tied to purchasing or messaging, the consequences leave the home entirely.

Studying it

Two established lines. One measures false wakes and misrecognition: replay TV programmes and everyday conversation corpora in controlled acoustic settings and in real homes, counting false wakes per hour and misexecution rates; the degradation of word error rate under noise and far-field conditions is a standard result and serves as a baseline for entry quality. The other is field research on household voice use: the distribution of false wakes, which devices households dare control by voice, and how users attribute and repair misexecutions (predominantly interview-based; stated generically).

Typical variables: wake threshold, presence or absence of a confirmation step before execution, impact level of the device. One methodological caution: false-wake rates measured in quiet labs drastically underestimate real homes — TVs, children, and multi-room sound sources are the main producers, so measurement must use real acoustic scenes or equivalent corpora.

Where it stops holding

• Not every misrecognition should be intercepted. Confirming every command strips voice of its hands-free value. Intercept by consequence tier: low-stakes devices (lights, music, blinds) execute directly and absorb occasional errors; high-stakes ones (locks, payments, outbound messages) require confirmation or lose voice access entirely.
• False-wake rates vary enormously across home acoustics. A quiet solo household and a TV-always-on family household can differ by an order of magnitude; one factory setting cannot fit both — sensitivity must be adjustable.
• Children's voices are the weak zone. Acoustic features of children raise recognition error rates, and children are precisely the keenest commanders of speakers; safety-relevant devices need child-speech scenarios in their test sets.

Applying it

• High-impact devices default to voice read-only: voice may query state; execution goes through the app or a physical entry. Where voice execution is genuinely needed, add one confirming step (echo back the intended action).
• Give every voice execution an undo window: "Living-room light off — say 'undo' to restore", cutting the repair cost of a misexecution to a single phrase.
• Make wake sensitivity user-adjustable, with a feedback affordance for false wakes ("you weren't being addressed") feeding personalisation.
• Provide a physical mute or timed pause for TV-in-the-room scenarios.
• How to check: replay a corpus of TV dialogue and chatter, measuring false wakes and misexecutions per hour; then audit the voice permission matrix per device — for each device answer "what is the worst outcome of one misrecognition?", and downgrade anything unanswerable or catastrophic.

Related

• Same group: Z4.10.1 Voice, app, and physical control of one device need real-time state sync · Z4.10.3 The directness of physical switches makes them the reliable fallback in failures · Z4.10.4 Users need to know the response-latency differences between entries
• Nearby: Z2.02 Uncertainty of inference · Z4.09.4 After recovery, state must be confirmed, not assumed restored
• Search terms: misrecognition · false wake word · unintended activation · word error rate