Cantonese Meeting Transcription: The Complete Guide Why Cantonese is hard for AI, and the engineering decisions that make accurate, decision-grade transcripts work.

Cantonese is spoken by an estimated 85 million people worldwide — across Hong Kong, Macau, Guangdong, and a global diaspora — making it one of the world’s most widely spoken languages. Yet it has been chronically underserved by transcription tools. The reasons are partly technical: six tones, a colloquial-versus-written gap that English has no analogue for, and a population that switches mid-sentence into English and Mandarin without thinking about it. The reasons are also commercial — most transcription R&D is concentrated on languages with a larger global content footprint, so off-the-shelf models treat Cantonese as an afterthought.

This guide walks through what we’ve learned building Oak’s Cantonese-first stack, and what it takes to get accurate, decision-grade transcripts out of the meetings Hong Kong teams actually run — from the engine that handles code-switching to the way a transcript becomes evidence.

What’s in this guide

This is the pillar overview. Each chapter below tackles one slice of the Cantonese transcription problem in depth, and links back here. Read them in order for the full picture, or jump to the one that matches the problem in front of you.

• How Oak Handles Cantonese–English Code-Switching: the token-level decisions a model has to make when a single sentence switches between Cantonese and English, why detector-based engines fail at the switch, and worked examples from real product, sales, and finance meetings.
• Cantonese Business Jargon: Keeping Nuance in Transcripts: the colloquial-versus-written gap, the near-homophones that flip meaning in finance and legal contexts, and how a Jargon Library keeps a team’s vocabulary correct.
• From Transcript to Meeting Minutes and Action Items: the four-step pattern that turns raw transcript into a decision-ready record, and how Oak structures every set of minutes.
• Timestamped Transcripts and Speaker Identification Explained: how diarisation separates who-said-what, why most engines lose speaker identity, and what makes a timestamp citation-grade.

The hidden cost of a bad Cantonese transcript

The case for getting Cantonese transcription right is easy to underestimate, because the cost of getting it wrong is spread out and rarely measured. It shows up first as the time someone spends cleaning up a machine transcript by hand — often the same junior staffer who was supposed to be freed up by automation in the first place. It shows up again as the meaning that quietly goes missing: the Cantonese aside that explained why a customer is unhappy, the code-switched commitment that never made it into the English summary, the flipped finance term that turned a loss into a neutral statement. None of these are dramatic on their own, but together they mean two halves of a company end up working from different versions of the same conversation.

For regulated teams the cost is sharper still. A transcript that quietly translates colloquial Cantonese into formal Chinese has made an editorial choice that may be disputed later; a transcript that cannot point to the exact second behind a contested line is not much use in an audit. The teams that care most about Cantonese transcription — broadcasters, law firms, statutory bodies — are precisely the ones for whom an inaccurate or unverifiable record carries real consequences. That is why this guide treats accuracy, attribution, and verifiability as a single standard rather than three nice-to-haves.

Why is Cantonese so hard for AI to transcribe?

Cantonese is tonal, colloquial, and rarely written the way it is spoken. A speaker says 嗰個 and means “that one.” A sales presenter says 我哋有個 quarterly target 要 hit and a Cantonese-only model breaks halfway through the sentence. An interview subject uses 蝕 versus 攝 — one means lose money, one means slot in — and a context-blind model picks the wrong one. Multiply this by 60 minutes of meeting and three speakers, and you have a transcript no one can use.

Three forces compound, and each chapter of this guide tackles one of them:

• Tone and acoustics. Six tones mean small pitch differences carry whole meanings, and near-homophones are everywhere — which is why a context-aware engine and a Jargon Library matter so much.
• The colloquial-vs-written gap. Hong Kong professionals speak in colloquial Cantonese but write in formal Chinese — so “accurate” transcription is also a translation choice. Covered in business jargon.
• Code-switching. English and Mandarin tokens land inside Cantonese sentences without warning. Covered in code-switching.

Tones and near-homophones

Cantonese is a tonal language with six tones, which means the same syllable carries entirely different meanings depending on pitch. For a listener this is effortless; for a model trained mostly on non-tonal languages it is a constant source of error, because the acoustic difference between two words can be smaller than the background noise in a typical meeting room. The result is a transcript littered with near-homophone substitutions — plausible characters that are simply not the ones the speaker said. In casual conversation this rarely matters, but in a board meeting or a legal proceeding a single flipped character can invert the meaning of a sentence. This is why a Cantonese-first engine has to combine acoustic modelling with context: weighing the surrounding words before committing to a character, rather than picking the most statistically common option in isolation.

The colloquial-versus-written gap

English has nothing quite like the gap between spoken and written Cantonese. Hong Kong professionals speak in colloquial Cantonese — full of particles, contractions, and turns of phrase that are rarely written down — but when they produce a formal record, they write in standard written Chinese. A transcript therefore has two legitimate targets, and a generic model usually lands awkwardly between them: too colloquial to file as a formal record, too formalised to capture what was actually said. Oak’s approach is to preserve the colloquial form faithfully and offer a clearly-labelled written version when one is needed, so the team decides which register the record should take rather than having the tool silently choose for them.

Code-switching as the default register

The third force is the one outsiders underestimate most. In Hong Kong workplaces, mixing English words into Cantonese sentences is not a quirk or a sign of imperfect fluency — it is the standard professional register. Technical nouns, product names, and business verbs arrive in English while the grammatical frame stays Cantonese, and speakers switch back and forth several times in a single sentence without thinking. Any engine that treats one language as primary and the other as an exception will break at every switch. The only robust approach is to treat Cantonese and English as a single space from the start, which is exactly what the code-switching chapter explains in detail.

Generic transcription vs a Cantonese-first engine

Most teams arrive here after trying a general-purpose tool and getting transcripts they had to clean by hand. The difference is architectural, not cosmetic:

How the pieces fit together

The reason these capabilities are documented as one guide rather than five separate features is that they only work as a stack. Accurate Cantonese transcription is the foundation: get the words wrong and everything built on top inherits the error. The single bilingual model produces clean text through code-switches; the Jargon Library locks in the names and terms that a general model would flatten; diarisation attributes each line to the right speaker; and second-level timestamps anchor every line to the audio. Only once all four are in place does the final step — turning that transcript into minutes and action items — become trustworthy.

Put the other way round, a weakness anywhere upstream shows up downstream. A transcript that mangles a code-switched sentence produces a summary that omits the decision buried in it. A transcript that flips a finance term produces minutes that record the opposite of what was agreed. A transcript that loses a speaker’s identity produces an attribution dispute no one can resolve. This is why the guide treats the problem holistically: a team evaluating a Cantonese transcription tool should test the whole chain on their own audio, not just glance at a headline accuracy figure.

What “decision-grade” actually means

Throughout this guide we describe the goal as a decision-grade transcript, and it is worth being precise about what that means. A decision-grade transcript is one a team can act on without going back to the recording to check. In practice that requires three properties at once. It has to be accurate — the words are right, including the jargon and the code-switched terms. It has to be attributable — every line is tied to the person who said it, consistently across the meeting. And it has to be verifiable — any line can be traced back to the exact moment in the audio in seconds, so a dispute is settled by evidence rather than memory. A transcript that has only the first property is a readable record; a transcript that has all three is something you can base a decision, a deliverable, or a legal position on. Everything in this guide is in service of that bar.

Who this guide is for

Hong Kong operators running bilingual meetings who are tired of cleaning up bad transcripts. Procurement and IT teams comparing AI transcription vendors. Editorial and post-production staff working with Cantonese audio at scale. Legal, compliance, and statutory teams for whom the transcript is a record, not a convenience.

Each article links back to the relevant Oak solution page, so you can see the engineering deployed against a real customer problem — broadcasters, law firms, statutory bodies — instead of as abstract feature claims. If you are comparing vendors, the most useful thing you can do is take a real recording of one of your own meetings, with its code-switching and its in-house vocabulary intact, and run it through each tool you are considering. The differences this guide describes are obvious the moment you see them on audio you actually recognise.

What changes once the transcript is reliable

When a team stops fighting its transcripts, the downstream effects compound quickly. The most immediate is that the minute-taker gets their job back: instead of half-listening and half-typing through a meeting, they participate fully and spend a couple of minutes afterwards confirming a draft. The quality of the meeting itself improves, because the person who used to be buried in note-taking is now in the conversation. The second effect is that the record becomes something people actually use — searchable across the workspace, consistent in structure, and trustworthy enough that a colleague who missed the meeting reads the summary instead of asking for a recap. The third, slower effect is institutional: a backlog of reliable, attributable transcripts becomes a genuine knowledge base, one a team can search months later to settle exactly what was agreed and by whom.

None of this happens if the underlying transcript is unreliable. A searchable archive of inaccurate transcripts is worse than no archive, because it gives false confidence. This is the throughline of the entire guide: the value is not in the novelty of AI transcription but in whether the output is good enough to build on. The chapters that follow each address one reason a Cantonese transcript might not be — and what it takes to fix it.

Frequently asked questions