Dialects share vocabulary with the standard but differ in pronunciation, so they are misread as near-homophones

What it is

Many Sinitic dialects share most of their vocabulary and the intended Chinese characters with Standard Mandarin, but initials, finals, and tone rules differ. A recognizer that aligns those words only to Mandarin phonology maps dialect pronunciations onto different characters that are close in the standard. The speaker said that lexeme; the screen shows a near-sound typo. This is not a foreign insert, and not a missing word in the lexicon. It is the same lexeme through another pronunciation rule.

Why it happens

An acoustic model trained on the standard learns “this initial + this final → this character.” A dialect treats historical voiced initials, entering tones, or nasal codas as a different set of contrasts, so the same lexeme lands in another cell of standard space. The decoder then picks the nearest standard character: one word becomes a predictable neighbour. If the language model wants fluent Mandarin, it may pull the whole sentence toward another fluent sentence that was not meant. Accent usually means offset inside the standard; dialect phonology can change the contrasts themselves, so the confusable set is larger. Because the target orthography is still those characters, the errors look like homophone substitutions; the root is phonological mismatch.

Studying it

Have standard and dialect speakers read the same lexemes, force a Mandarin decoder, and test whether substitutions fall in a predictable near-character set. Separate “lexeme right, character wrong” from “lexeme recognized as a different word.” The control is the same speaker reading again in Mandarin. Low error on the Mandarin pass and stable neighbour substitutions on the dialect pass isolate phonological rules from generic accent noise. Do not score Mandarin output against a dialect-pinyin reference; that entangles orthography with recognition.

Where it stops holding

Languages with a separate written tradition and little shared lexicon are not “shared vocabulary, different pronunciation,” and need another recognizer. Contact varieties already shifting toward Mandarin have a smaller confusable set. If the output target is dialect orthography or phonetic spelling rather than standard characters, the task becomes a different transcription problem. Noise and channel can also create near-sound swaps; confirm regular confusions on clean recordings first.

Applying it

• For a target region, collect a dialect-to-standard confusion table and turn stable high-frequency neighbours into proofreading candidates rather than only red marks.
• Let people declare a regional variety and load the matching pronunciation lexicon; do not imply that “speaking slowly will become Mandarin.”
• Accept with matched dialect/Mandarin pairs of the same lexemes, and report the rate of predictable neighbour substitutions rather than only overall WER.

Related

• Same group: C7.04.1 Training-data distribution determines whose speech is recognized well · C7.04.2 Recognition degrades on mixed Chinese–English and dialect words · C7.04.3 Accuracy gaps exclude particular populations · C7.04.4 Mid-utterance language switches require live language identification · C7.04.5 A wrong language decision forces the rest of the sentence through the wrong phonology · C7.04.7 Accent and dialect gains depend on training data from those speakers, not on algorithms alone
• Adjacent: C7.03 Types of Recognition Errors · C7.13 Voice Editing and Spoken Correction
• Search: dialect phonology · near-homophone substitution · topolect ASR