Cantonese business jargon is hard to transcribe because professionals speak colloquial Cantonese but write formal Chinese, and many terms are near-homophones — words that sound alike but mean opposite things, so a generic model flips the meaning. The fix is a Jargon Library, applied per project and owned in your workspace, seeding the model with your terms.
Every industry has jargon, but Cantonese business jargon comes with an extra layer: a written–spoken gap. Bankers, lawyers, and broadcast presenters speak in colloquial Cantonese yet write in formal Chinese — and the assumption that a transcript should be one or the other gets you a transcript that is neither. The terms that carry the most meaning are often the ones a general-purpose model is least equipped to hear.
Key takeaways
- The written–spoken gap is the root problem. Speech is colloquial Cantonese; the record is expected in formal Chinese — so transcription is also a translation decision.
- Near-homophones flip meaning. Terms like 蝕 (lose money) versus a similar-sounding character can invert a sentence; in finance and law, that is the difference between right and wrong, not rough and polished.
- It is a failure of meaning, not spelling. The model transcribes a plausible character — it just isn't the one the speaker meant.
- A Jargon Library is the fix. Seeding the model with your own vocabulary — applied per project, owned in your workspace — resolves most ambiguities at first pass and keeps your terms private to you.
- Libraries need curation. Seed the obvious, add terms as they're missed, and review monthly — jargon evolves, and an un-curated library decays.
Why is Cantonese business jargon hard to transcribe?
Two forces combine. First, the colloquial-versus-written gap: a Hong Kong banker speaks one way and files minutes another, so “accurate” has two possible targets and a generic model often lands between them. Second, Cantonese is dense with near-homophones — words separated by a single tone or a subtle vowel that mean entirely different things. In casual speech the listener disambiguates from context. A context-blind model can’t, so it gambles on the most statistically common character and loses.
Where nuance disappears
A finance director says 個 cap table 蝕緊本 — the position is losing money. A generic model transcribes 蝕 as a similar-sounding character, and the meaning flips to its opposite. A legal partner uses 單嘢 to refer to a specific case under privilege; a model that doesn’t know the context renders it as a generic “thing.” These are not failures of literal transcription — they are failures of meaning, and they are invisible until someone downstream acts on the wrong word.
Jargon that trips up generic models
A few illustrative examples of how meaning slips, by sector. These are representative of the patterns we see, not an exhaustive list:
| Sector | Spoken term | Intended meaning | How a generic model can mishear it |
|---|---|---|---|
| Finance | 蝕緊本 | Currently losing the principal / underwater on the position | Swaps 蝕 for a near-homophone, flipping loss into something neutral |
| Finance | cap table | The capitalisation table — who owns what equity | Phonetically mangled into unrelated Cantonese characters |
| Legal | 單嘢 | This particular case / matter (often under privilege) | Rendered as a generic "thing," stripping the legal reference |
| Legal | 過數 | To transfer funds / settle a payment | Split or mis-segmented, losing the financial action |
| Broadcast / TV | show-specific names & segment titles | Programme, segment, and on-air talent names | Treated as common words, so a usable summary needs a full rewrite |
The pattern is consistent: the words that matter most to the record are the ones a model trained on general Cantonese is least likely to get right.
How Oak preserves it
Jargon Libraries let teams seed the model with their own vocabulary — names, products, regulatory terms, internal abbreviations. A library is applied per project and owned in your own workspace, so a law firm’s privileged terms stay under your control and never become another client’s. Combined with context-based decoding — where the model weighs surrounding words before committing to a character — this resolves most jargon ambiguities at first pass. For the full mechanics of setting one up, see building a Jargon Library.
Patterns that work
We see three patterns in customer libraries:
- Seed with the obvious. Company name, product names, and the top 20 internal acronyms. This catches the majority of mistranscriptions on day one.
- Add over time. When a transcript misrenders a recurring term, add it once — and Oak corrects it from then on.
- Review monthly. Jargon evolves; libraries that don’t get curated decay. A short monthly pass keeps it sharp.
Where this shows up
Most visibly in Oak for Legal Sector — where preserving the precise legal phrasing is the whole job — and across the TV Channels deployment, where show-specific terminology is the difference between a usable summary and one that has to be rewritten. The same Jargon Library mechanism keeps a team’s domain vocabulary accurate on their specific audio over time.
Oak for Legal Sector
The customer-facing deployment that uses the workflow described in this article.
Frequently asked questions
Why does generic AI mistranscribe Cantonese business terms?
Two reasons. Hong Kong professionals speak colloquial Cantonese but write formal Chinese, so a transcript has two possible targets and a generic model lands between them. And Cantonese is full of near-homophones — words a single tone apart that mean opposite things — so a context-blind model gambles on the most common character and often loses. The error is a failure of meaning, not spelling.
What is a Jargon Library and how does it help?
It is a list of your team's own vocabulary — company and product names, acronyms, regulatory terms, and internal jargon — that the model uses when transcribing. You apply a library per project, and it's owned in your own workspace. Combined with context-based decoding, it resolves most jargon ambiguities on the first pass, so recurring terms come out right every time instead of only when the audio is clean.
Will my company's confidential terms stay private?
Yes. A Jargon Library is owned in your own workspace, so a law firm's privileged terms or a media company's unreleased show names stay under your control and never become anyone else's. Applying a library per project also keeps a sensitive matter's vocabulary scoped to that matter. This is why legal and statutory deployments rely on the feature.
How do I keep a Jargon Library accurate over time?
Three habits. Seed it with the obvious 50 entries on day one (company name, top products, top acronyms). Add any term the moment a transcript misrenders it. And review it monthly, because jargon changes and an un-curated library decays. See building a Jargon Library for the full workflow.
Does this matter for legal and broadcast transcripts specifically?
Especially there. In legal work the precise phrasing is the record, so a flipped term can mislead a matter. In broadcast, show- and segment-specific names appear constantly, and getting them wrong turns an automatic summary into a manual rewrite. Both rely on the Jargon Library plus context-based decoding to hold nuance.