Why Your Cantonese Meeting Notes Always Feel Slightly Off: Two Hidden Fault Lines

Most people react the same way the first time they use an AI tool to take notes on a Cantonese meeting: it looks credible, but something feels off. There's no obvious error on the page, yet compared against what you remember of the conversation, something has gone missing.

That gap usually isn't the thing people instinctively blame — raw recognition accuracy. For pure Cantonese, a single speaker, and a quiet room, today's mainstream tools are genuinely usable. The trouble lies in two other places, and they happen to be the two most common shapes of everyday conversation in Hong Kong.

First, recognition quality drops sharply the moment a single sentence mixes Chinese and English. Second, even when the system hears you correctly, the text it outputs has often already been rewritten into Standard Written Chinese — no longer the sentence you actually said.

These two problems have different causes and affect different situations, so they're worth separating.

The 30-second version

• Code-switching has a cost, and that cost is measurable → Published research reports character error rates of roughly 20%–26% for Cantonese-English code-switched content, versus roughly 6%–8% for comparable systems on monolingual content. A three- to four-fold gap.
• "Heard correctly" and "written correctly" are two different things → You say "幾時搞掂" (when will this be done?); the notes say "什麼時候完成" — the meaning survives, the actual words don't.
• This is irrelevant in most settings and critical in a few → Casual meeting summaries are unaffected. Medical consultations, social work case notes, mediation records, and research interviews are the opposite.
• Cantonese data volume changed by an order of magnitude in 2025 → From public corpora measured in dozens of hours to WenetSpeech-Yue's 21,800 hours — but code-switching remains the unsolved part.
• Before choosing a tool, decide whether you want the meaning or the actual words → The two need different settings. There's a checklist at the end.

First, separate two things: recognition vs. presentation

When people discuss Cantonese transcription quality, two independent stages get conflated.

Recognition (ASR) turns sound into text. Failure here means mishearing — the system can't tell what you said, or substitutes a similar-sounding word.

Presentation decides what form that text finally takes. Failure here means rewriting — the system heard you perfectly, then smoothed your colloquial speech into formal written language.

Most complaints about "inaccurate Cantonese notes" are actually these two layers failing separately, and they need completely different fixes. The first is a question of model capability and training data. The second is usually a design choice — and often a deliberate one.

Fault line one: Cantonese-English code-switching

Why this is unusually hard

「我想去 Causeway Bay 開個 meeting，順便 check 一下 schedule。」 (Gloss: "I want to go to Causeway Bay for a meeting, and check the schedule while I'm at it." — the English words appear mid-sentence in an otherwise Cantonese utterance.)

Nothing about that sentence is remarkable in Hong Kong, but it demands something entirely different from a recognition system than a monolingual Cantonese sentence does. The system has to switch between two completely different phonological systems and vocabularies within a single sentence — sometimes within a single breath — with no advance warning of where the switch happens.

Worse, when Hong Kong speakers use English words, the pronunciation has usually already been Cantonesized: syllable structure and tonal contour have both been localized. These pronunciations neither match what an English model expects nor appear in a Cantonese model's lexicon. Both sides treat them as foreign.

The gap in numbers

Published research paints a consistent picture: character error rates for Cantonese-English code-switched content land in roughly the 20% to 26% range, while comparable systems on monolingual content sit around 6% to 8%. On dedicated Cantonese-English mixed speech benchmarks, Whisper-based systems show mixed error rates of about 22%.

In practical terms: roughly one error every ten-plus characters for pure Cantonese, versus roughly one error every four to five characters for code-switched content. The latter is enough that a transcript needs line-by-line verification before you can rely on it.

These errors have an unusual shape

Worth noting: code-switching errors don't look wrong the way ordinary recognition errors do. The system rarely leaves a gap or emits garbage. It tends to guess a Chinese word that sounds close and is grammatically plausible, and slot that in.

The result is a sentence that reads perfectly smoothly while the meaning has already drifted. That's far more dangerous than an obvious typo — an obvious error makes you stop and check; a fluent wrong sentence you simply read past.

Fault line two: spoken Cantonese vs. Standard Written Chinese

The same sentence, two ways of writing it

Nothing in the middle column is wrong. The meanings are accurate and it arguably reads more smoothly. The problem is that nobody talks like that.

Why systems do this

This is usually not a bug. Training data for written Chinese is overwhelmingly Standard Written Chinese; written colloquial Cantonese (嘅, 咗, 喺, 哋) is a vanishingly small share of the overall corpus. Absent an explicit instruction, a model naturally outputs the variety of Chinese it knows best.

And in most cases that default is right. For meeting summaries, action items, or anything sent to colleagues who don't speak Cantonese, the written-Chinese version is clearly more useful.

But some situations are exactly the opposite

The trouble is that an entire class of use cases needs precisely the original words.

• Medical consultations: the words a patient uses to describe a symptom are themselves clinical information. Once 「頂住頂住」 (a persistent pressing/blocked sensation), 「唧住痛」 (a squeezing pain), or 「作嘔作悶」 (queasy and nauseated) get rewritten into standard terminology, that specific texture is gone.
• Social work case notes: a service user's own words feed into case assessment; rewriting adds a layer of interpretation at the recording stage.
• Mediation and dispute handling: who said what, and in what tone, is often the substance of the dispute itself.
• Research interviews and oral history: the value of a verbatim transcript rests on it being verbatim.
• Language teaching and corpus building: spoken forms are the object of study.

What these share is that the record isn't just there to remind people what was discussed — it may later be quoted, reviewed, or treated as evidence. A quietly tidied-up transcript is risky in that role, precisely because it looks entirely normal. You have no way of knowing what was changed.

An easily missed precondition: the state of Cantonese data

To understand why Cantonese tools perform so unevenly, it helps to look at the scale of available data.

The arrival of WenetSpeech-Yue in 2025 was an order-of-magnitude shift. Before it, public Cantonese speech data had long been stuck at the tens-to-hundreds-of-hours level, while major languages routinely count in the tens of thousands. That explains why Cantonese tools have generally lagged.

But note what more data mainly improves: monolingual Cantonese recognition. Code-switching needs mixed-language corpora, which are harder to annotate and still thin on the ground. So fault line one will improve noticeably more slowly than pure Cantonese recognition does.

The wider Hong Kong context

One related set of figures is worth mentioning, because it shapes what a "Cantonese tool" actually has to handle. According to the 2021 census, Hong Kong has roughly 619,568 non-Chinese residents, or 8.4% of the population. On Cantonese speaking ability: 17.6% among Filipino residents, 25.0% among Indian residents, and 34.7% across South Asian residents overall — while English ability generally exceeds 84%.

In other words, everyday communication in Hong Kong isn't a single-language Cantonese setting. Cantonese, English, and several other languages are present at once. A tool that recognizes Cantonese well but can't translate in real time solves only half the problem — and the same holds for any multilingual city.

How Traverba handles these two fault lines

• Code-switching first — recognition is tuned for Hong Kong-style mixed Chinese-English sentences, rather than treating Cantonese and English as two modes you must choose between.
• Spoken form preserved — transcription follows the form you actually spoke; summaries and translations are generated separately. The original words and the tidied version coexist, rather than the tidied version overwriting the original.
• Real-time transcription and translation in one — one conversation, one speaker, 100+ languages. Audience members scan a QR code to read subtitles in their own language, with no app download.
• On-device by default — speech recognition runs on the device by default, with cloud AI as an option. For sensitive content, nothing has to leave the device.
• Full verbatim transcript export — after the session you can export both the verbatim original and the translated version, stored separately.

To be clear: if your use case is ordinary business meeting summaries, internal sync-ups, or action item tracking, most AI note tools on the market already do this well, and written-Chinese output is arguably more useful. There's no need to switch tools for this. These differences only matter where the original words are themselves the data.

How to decide: a short checklist

Ask yourself three questions.

1. Will this record later be quoted or reviewed? If yes (case notes, clinical records, mediation, interviews), you need a verbatim transcript, and you need to confirm the tool isn't rewriting behind your back. If no, a written-language summary is more practical.

2. How dense is the Chinese-English mixing? An occasional English word is something most tools handle. If it's the typical Hong Kong conversation that switches two or three times per sentence, test with a real recording first — don't rely on the official language list.

3. Is the content sensitive? For recordings involving medical records, casework, or legal matters, confirm where the data goes: processed locally or in the cloud, retained for how long, used for training or not.

One final piece of advice: whichever tool you pick, test it with a real conversation of your own, not a read-aloud script. Read-aloud Cantonese has no code-switching, no colloquial particles, and no overlapping interruptions — it won't surface any of the problems above.

Learn more

There's no single tool that suits every Cantonese transcription situation. The key is deciding first whether you want the meaning or the actual words — that decision directly determines what to choose and how to configure it.

To see how Traverba handles code-switching and spoken-form preservation, visit traverba.com, or download it on Google Play or the App Store.

Error rate figures cited here come from published Cantonese speech recognition research; corpus scale figures come from official publications by each corpus; population and language ability figures come from the Hong Kong Government's 2021 Population Census. All data is current as of mid-2026. Model and tool capabilities change continuously, and real-world performance varies with recording quality, accent, ambient noise, and specific version. All brands and trademarks belong to their respective owners.