A wrong language decision forces the rest of the sentence through the wrong phonology

What it is

Once language identification labels the current stretch as language A, later frames follow A’s phoneme inventory, pronunciation lexicon, and script. If the speaker has already switched to language B, the rest of the sentence is still decoded through A’s phonology: English is broken into Chinese characters, or Chinese is spelled as a Latin-letter string. The output often looks like garbage or bizarre homophones, not a simple typo. The failure is a locked route, not a missing loanword in the lexicon.

Why it happens

After a language is chosen, the decoding graph prunes other languages to save compute. Acoustic features map onto that language’s emission model; B’s phones can only project onto the nearest sequence in A. A Chinese-to-English lock-in turns a Mandarin tail into pinyin debris or English neighbours; an English-to-Chinese lock-in yields characters unrelated to the source. Streaming partials show the mismatch early; if the final pass never re-runs LID, the mismatch is committed. Confidence can still be high—the match is internally consistent inside the wrong phonology. This is not a homophone substitution, which picks the wrong item inside the correct lexicon; here the lexicon itself is the wrong one.

Studying it

At known switch points, inject a wrong language tag, or use an LID that fails there, and compare transcripts of the tail under correct versus locked labels. Record which writing system the mismatch emits, whether people can tell from text that “the language is wrong,” and whether they respeak the whole sentence. Independent variables include switch location, phonological distance, and whether the tag can be revised mid-stream. Do not report only LID accuracy: accuracy can be high and still lock the tail at the critical switch.

Where it stops holding

If the utterance never switched, there is no forced tail; that is ordinary monolingual error. Systems that keep multilingual n-best paths can revise at the end, so lock-in is not absolute. A very short tail (one English abbreviation) can look like a mixed-lexicon miss; separate it by whether the phonology changed. A foreign-accented speaker of language A may make LID chatter while decoding stays inside A, which is not this lock-in.

Applying it

• When partials suddenly change writing system, or a stretch becomes unsegmentable characters, offer “decode again in the other language” rather than only respeak.
• If the language posterior flips mid-sentence before commit, re-decode the tail; do not execute the locked hypothesis.
• Put “first half right, second half in the wrong script” in acceptance tests, and confirm recovery without repeating the first half.

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 · C7.04.6 Dialects share vocabulary with the standard but differ in pronunciation, so they are misread as near-homophones · C7.04.7 Accent and dialect gains depend on training data from those speakers, not on algorithms alone
• Adjacent: C7.10 Live Partial Results · C7.03 Types of Recognition Errors
• Search: language identification error · phonological mismatch · forced decoding