Recognition degrades on mixed Chinese–English and dialect words

What it is

Everyday speech often drops an English word into a Chinese sentence, or folds a dialect-specific lexical item into a Mandarin frame. Code-mixing and dialect words are out-of-distribution for a recognizer trained on one standard: English fragments are forced into Chinese syllables, dialect words are rewritten as near Mandarin neighbours. The degradation sits at the word: the mixed unit collapses. That is not the same as the whole sentence changing language.

Why it happens

A monolingual decoding graph’s lexicon and language model barely contain strings such as “打开 flight 模式” or a dialect pronoun inside otherwise Mandarin syntax. English phones run through a Chinese phonology and emerge as characters; a dialect word not in the lexicon can only land on a Mandarin neighbour. Context around the mix is also low-probability for the language model, so neighbouring standard words sometimes get dragged with it. This is not accent (the same standard word, pronounced differently): the missing piece is the word itself. In ads, product names, programmer talk, and support calls, such mixing is the regime, not noise.

Studying it

Build four scripted conditions—pure Mandarin, one English proper name inserted, a dialect word inserted, fully mixed sentences—and compare WER with keyword recall. English inserts should cover letters, digits, and common loans; dialect words should come from real regional use, not researcher-invented tags. Report whether the mixed fragment is preserved, not only whole-sentence WER: the sentence can look fine while the English product name is gone. News-read baselines will deny the mixing problem.

Where it stops holding

Systems with bilingual lexicons and mixed language models can be stable on frequent loans and still fail on user-invented English abbreviations. A whole sentence in dialect is closer to recognizing another language than to “one word inside Mandarin.” Users who only type formal language and never mix will not see this degradation. Blaming every Chinese–English error on a wrong language-ID decision misses the case where the English word was never in the lexicon.

Applying it

• Put product names, English command aliases, and high-frequency regional dialect words in the lexicon, and regress on mixed sentences rather than only pure Mandarin scripts.
• For English fragments that get swallowed into characters, let people type the proper name rather than forcing a full respeak.
• Before launch, collect real mixed speech in target cities and accept on “is the mixed fragment still there,” not only on overall character accuracy.

Related

• Same group: C7.04.1 Training-data distribution determines whose speech is recognized well · 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 · C7.04.6 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: C7.03 Types of Recognition Errors · C7.14 Command Grammar versus Free Expression
• Search: code-mixing · dialect word · mixed-language ASR