Correction can cost more than repeating the utterance

What it is

After a recognition error, two recovery paths are open: edit the existing text, or respeak the whole utterance. Editing means locate, select, replace; respeaking means assembling another breath and waiting through another decode. When the edit steps are slower and more error-prone than saying it again, correction costs more than a respeak. The claim is about the economics of recovering from recognition failure, not about how spoken commands edit a document.

Why it happens

A dictation hypothesis is linear text. Changing one word first requires finding it: placing a caret precisely on a touchscreen is expensive, and saying “the third character” is unstable. A local replacement that still goes through recognition may damage neighbouring words that were already right. Respeak discards the current hypothesis and decodes a new full utterance, using whole-sentence context, which is sometimes more accurate than patching a point. The cost structure depends on the channel: with a keyboard, changing one character is cheap; in eyes-free or driving use, any pointing action is dear and a full respeak is cheaper. The more scattered the errors, the closer total edit time comes to—and then exceeds—saying it once more. People also treat “try again” as teaching the system, even though the second attempt is not guaranteed to be better.

Studying it

Seed controlled errors in dictation and command tasks (substituted names, inserted function words, deleted negations) and offer edit / respeak / cancel. Measure completion time, key or touch count, whether the second attempt is still wrong, and workload. Independent variables include number and position of errors, presence of a keyboard, and eyes-free use. The comparison is recovery-strategy cost, not whether a particular spoken-edit grammar is usable. Offering only edit, with no respeak, yields coerced edit times and hides spontaneous choice.

Where it stops holding

A single homophone swap with a candidate bar is often cheaper to tap than to respeak the sentence. In long text with one error, respeaking throws away a mostly correct draft, so editing wins. When the network is poor and a second recognition is equally unstable, respeak becomes spinning. For speakers who stutter or have impaired articulation, respeak itself is costly; “just say it again” is not the cheaper default. The comparison also does not apply to commands already executed that text editing cannot undo—that is rollback, not changing characters.

Applying it

• After recognition failure, offer both respeak and edit; do not lock people into multi-step pointing by default.
• When errors exceed one or two spots, or no keyboard is present, make respeak the primary action and edit secondary.
• Accept recovery paths on task time: from noticing the error to submitting a correct result, respeak should not be so slow that nobody uses it, and edit should not be so buried that it might as well not exist.

Related

• Same group: C7.03.1 Substitutions, insertions, and deletions have different consequences · C7.03.2 Homophone errors cannot be detected from the text itself
• Adjacent: C7.13 Voice Editing and Spoken Correction · C7.17 Switching Between Voice and Keyboard
• Search: respeak · correction cost · ASR error recovery