Recognition results must be correctable
Aliases: speech correction · n-best TV search
What it is
Far-field mishearing is the default. The system has to put “what it thought you said” on the big screen and let the user change the wrong stretch—not respeak the whole sentence, and not walk an on-screen keyboard from the first letter. Recognition results must be correctable is the condition under which living-room voice search gets used a second time. If the only move after a miss is start over, people switch to cover browsing and the speech channel is abandoned.
Why it happens
One wrong character in a media query retrieves a different title. The miss is often part of a proper name: a surname, a numeral in the title, a slice of a foreign original. Respeaking the whole utterance gambles the already-correct parts again; under far-field interference the second try is not reliably better. Local correction turns the cost from “say it again” into “pick once among candidates” or “change that word.”
Fixing characters on TV is itself expensive, so the correction UI cannot be an ordinary field. What works is an n-best list (what else it might have been), word-level highlight (focus the wrong word, offer near-sound replacements), and “filter within these results” rather than clearing the search. The default action on OK must be “search the sentence I can see,” not jump into a wrong result page before the transcript has been read.
Invisible recognition is uncorrectable. Results without a transcript leave the user unable to tell mishearing from “not in the library,” so the next move is blind respeaking.
Where it stops holding
Short wake commands that hit first time (“pause,” “next episode”) do not need transcript-level correction; saying the command again is cheaper. When the utterance was heard as a wholly unrelated sentence, n-best will not contain the right one; local substitution cannot save it, and the exits are “say it again” and keyboard fallback. In kids or game-like settings, reading the wrong sentence aloud can embarrass; a silent correction list is better than TTS. If speech is the only input for an accessibility user, correction must also be speakable (“the third one,” “not that”) rather than a D-pad-only list.
Applying it
- Show the transcript on every query and wait for confirmation. Default focus on “search this,” with focusable alternative sentences beside it.
- Make a wrong word independently selectable and replaceable by a near-sound or by typing only that word. Do not open a full on-screen keyboard to fix one character.
- After two misses, change the exit to cover browsing or “type on the phone,” rather than a third “please say that again.”
- Verify with easily confused titles and queries that contain numerals or foreign words. After the first recognition miss, can the user correct in two steps without speaking again? For failures, record whether the transcript was invisible, the right candidate was missing, or correction itself required a keyboard.