Homophone errors cannot be detected from the text itself

What it is

A homophone error writes a word that sounds the same, or nearly the same, as what was said. The output is well-formed text and sounds right, so the transcript often does not reveal the mistake. Chinese characters that share a pronunciation, English pairs such as there/their, and collisions between names and common nouns all produce this class. It is a subclass of substitution whose damage is invisible to the user: no red underline or garbage string flags that something is wrong.

Why it happens

The acoustic model yields a neighbourhood of phones or syllables; which word is chosen depends on the language model, the domain lexicon, and contextual biasing. When context is weak—sentence-initial position, an isolated slot, a rare name—the decoder falls on the homophone that was more common in training. This is acute in Chinese: one pinyin string maps to many characters, and the language-model prior almost alone decides. People proofreading for “does this look fluent” will miss the swap; they have to compare against the intended pronunciation to notice that a meeting became a different character sequence with the same sound. Playback can expose the problem, but dictation by default delivers text, not audio. Confidence is often high on both homophones because the acoustics really do match.

Studying it

Build minimal pairs: same sound, different character, different slot consequence (name versus common noun, place versus verb). Ask people to find errors from text only, from playback only, and from text plus audio. Dependent measures include detection rate, time to detection, and rate of wrong execution. Report homophone substitutions separately from ordinary substitutions. Laboratory vocabularies of rare words overstate how easily people catch errors from text; high-frequency homophones are the everyday misses.

Where it stops holding

In strongly tonal or richly contextual sentences, homophones are suppressed and errors look more like non-homophonic acoustic confusions. Dialect and accent rewrite the homophone set: characters that merge for one group of speakers do not merge for another. Input methods that still show phonetic spelling let people check sound against character; plain-text dictation does not. Do not call every fluent-looking error a homophone: a wrong digit or a scrambled word order can also look fine, by a different mechanism.

Applying it

• For homophone-dense slots such as names, addresses, and account identifiers, require read-back, a candidate list, or contacts biasing rather than dropping a single line of text.
• Offer playback of the original utterance in the proofreading view so people can catch swaps by ear rather than by skimming.
• In acceptance tests, include cases that “look right and execute wrong,” and count homophone errors submitted without correction.

Related

• Same group: C7.03.1 Substitutions, insertions, and deletions have different consequences · C7.03.3 Correction can cost more than repeating the utterance
• Adjacent: C7.04 Accent, Dialect, and Code-Switching · C7.13 Voice Editing and Spoken Correction
• Search: homophone · contextual biasing · undetectable ASR error