When only a command grammar is supported and the phrasings are unstated, people trial-and-error

What it is

The product only accepts a command grammar, while the interface pretends “say anything.” People do not know the legal phrasings, so they reword until they happen to hit the graph, or they quit. This hidden grammar turns the learning burden into live trial-and-error, and the misses look like recognition inaccuracy.

Why it happens

The grammar constraint is hard on the decoder and open in the user’s model. If each attempt returns only “didn’t catch that” or an unrelated execution, the feedback contains no “shape of what you can say.” People try synonyms, politeness, slowing down—sometimes useful for open NLU, useless on a fixed graph, sometimes even further off. Trial-and-error burns patience and also ships error samples into logs that look like ASR problems. The discoverability failure is symbolic: the channel can hear, the words are not on the graph. This is not free-form ambiguity, where the sentence was understood and the wrong skill was chosen; here the sentence is not in the language at all.

Studying it

Give a task goal and no examples; log attempts until success or abandonment, how attempts drift from legal phrasings, and whether people think it was a recognition miss. Compare a condition that shows a legal example after failure. Gaze or spoken reports show whether they hunt through help. Internal staff as participants massively understate trial-and-error.

Where it stops holding

A handful of extremely natural phrases (“call mom”) hit often even if unpublished. If the grammar already overlaps colloquial speech, hiding it hurts less. A help document that exists but cannot be found during the task still counts as unstated. Treating trial-and-error as a need for a bigger model keeps the grammar hidden.

Applying it

• In the listening state, show one or two currently valid examples, and swap in another after failure, rather than only “please say that again.”
• List commands in settings or first-run, searchable.
• Split logs into “out-of-grammar utterances” and “in-grammar recognition errors”; the former is discoverability, the latter is ASR.

Related

• Same group: C7.14.1 Command grammars require fixed phrasings: more reliable recognition, at a memory cost · C7.14.2 Free expression lets people describe intent in natural language, at a large rise in parse ambiguity · C7.14.4 Command grammar and free expression can coexist in layers
• Adjacent: C7.16 Visible Feedback for Voice Input · C3.20 Gesture Discoverability
• Search: hidden grammar · trial-and-error · command discoverability