Spoken correction depends on saying a replacement command rather than editing text directly

What it is

Voice correction often does not move a caret onto the wrong character. It issues a replacement command: “change A to B,” “not …, … instead.” People edit by speaking the operation rather than by touching characters. If recovery from a recognition miss takes this path, the system must parse the utterance as an edit, not as new dictation.

Why it happens

A dictation channel’s default is to append every utterance as words. Correction has to encode a meta-operation on the same channel: the target span, the action (replace, delete, insert), and the new content. That needs a recognizable command pattern, or a parser that can pull edit intent out of natural language. The more fixed the pattern, the more stable and the less like talk; the freer it is, the more it collides with body text—“make it Friday” may be content or a command. The system must also align “A” to a span in existing text; a failed align edits the wrong place. This is not “correction can cost more than a respeak”: that compares recovery-path times. This describes how the spoken-replace path itself works.

Studying it

Give text with known errors and allow only spoken replace commands. Measure how often the command is treated as an edit rather than dictation, how often the right span is changed, and what phrasings people invent. Compare a prompted fixed pattern with no prompt. Include homophone targets and a word that occurs more than once. Do not test with a keyboard available; people will abandon the spoken path.

Where it stops holding

A candidate bar that can be tapped does not need a replace command. For a one-character miss, respeaking the sentence may still be simpler. Multilingual users may issue the command in a different language from the body, so the parser must first tell which layer is metalanguage. A pure dictation engine with no edit-intent classifier will write every correction utterance into the document.

Applying it

• Publish one or two discoverable replace patterns and prompt them with an example after an error, rather than assuming people will improvise the right sentence.
• Make “dictate / edit” an explicit mode, or confirm before applying when an edit verb is recognized.
• In acceptance tests with voice only, count replace commands swallowed into body text; that count should be near zero.

Related

• Same group: C7.13.2 The system must decide whether a correction targets recent text or an arbitrary earlier span · C7.13.3 Voice editing lacks the precise caret placement of keyboard editing · C7.13.4 Complex edits are less efficient to express in speech than by direct manipulation
• Adjacent: C7.03 Types of Recognition Errors · C7.14 Command Grammar versus Free Expression
• Search: spoken correction · change X to Y · voice editing