Transcription serviceCantonese

Cantonese Transcription Services

Cantonese transcription for Hong Kong audio, where the first question is not accuracy but which written system the transcript is in, because what a Cantonese speaker says and what Standard Written Chinese puts on the page are not the same words.

Free sample: Send a short Cantonese audio clip and we return a transcript in the format and orthography you specify, at no cost, before you commit to a project.

Specification

Audio
Wideband, narrowband, mono, or separated channels
Written form
Written Cantonese or Standard Written Chinese, picked before the first file and held for the whole batch
Characters
Traditional throughout, in Hong Kong forms, with full-width punctuation
Romanization
None by default. Jyutping or Yale is named in the scope when a romanized layer is ordered
Accuracy
Target agreed per project and measured on a held-out slice of your own recordings
Review
Second-pass review on a sampled percentage, rate agreed up front

Dialects and varieties covered

Hong Kong Cantonese.

Who does the work

Contributor floor
1,000+per market
Markets covered
Hong Kong
Dialects covered
1dialect

Coverage is drawn from Hong Kong, where Spirelight already publishes a delivered Cantonese dataset, so it is demonstrated rather than asserted.

Written Cantonese and Standard Written Chinese are two different deliveries

Cantonese is spoken in one system and written in another. A Hong Kong newsreader is reading Standard Written Chinese aloud in Cantonese pronunciation; two colleagues arguing about a delivery date are using their own vocabulary, particles and sentence patterns, none of which survive being written down as Standard Chinese. A transcript of the same clip can therefore be produced two ways, and the two files differ word by word rather than in style.

Written Cantonese also needs characters that do not occur in Standard Chinese text, among them the everyday negator, the copula, the possessive particle and the perfective marker Hong Kong speakers use in almost every sentence. Whether the transcript spells those out, or substitutes a Standard Chinese equivalent nobody said, is a convention rather than a preference, and a batch that answers it one way on Monday and another on Tuesday is no longer consistent with itself. The answer goes in the style guide before production, and Hong Kong delivery is in traditional characters.

Six tones, and a wrong one is still a real word

Cantonese separates six tones on the same syllable, and the contrasts are not academic. The syllable si carries poem, history, to try, time, market and to be, told apart by pitch alone. An annotator who hears the tone wrong writes a word that exists, fits the grammar and passes any spelling check, so tone accuracy here is a staffing question rather than a tooling one. The work is done by Hong Kong Cantonese speakers, and the review pass is run by people who speak the language rather than people who read it.

That same contrast is what makes romanized output worth deciding deliberately. Jyutping writes tone as a digit after the syllable, Yale uses diacritics and a silent letter, and much of what circulates informally is neither and marks no tone at all. Where a pipeline wants a romanized field for lexicon work or a pronunciation front end, the scheme is named in the scope, because batches romanized under different schemes cannot be merged.

English sits inside the sentence, not beside it

Hong Kong speech moves into English and back inside a single clause, and that is the ordinary way people talk rather than a lapse: an English noun, verb or piece of workplace vocabulary lands in a Cantonese frame, keeps Cantonese grammar around it, and often carries a locally settled pronunciation. A transcript that quietly translates those words, or drops them for being off-language, removes what made the recording Hong Kong audio.

So the scope states what happens to them: kept in Latin script, which is the usual answer, tagged for later filtering, or normalized, which is rarely what anyone wants once they see it. Numerals, dates, place names and company names get a stated form as well. All of it is settled before the first file, and the accuracy target is checked against a slice of your own recordings the production team never sees during the run.

This service is not the right fit when

  • Mandarin audio filed as Cantonese, or Cantonese audio filed as Mandarin. The two are staffed from different pools and quoted separately, and Mandarin work belongs on the Chinese transcription page in this cluster.
  • Taishanese, Hakka or Teochew audio. A Cantonese speaker does not follow them, so each is recruited, scoped and priced on its own.
  • One column holding written Cantonese in some files and Standard Written Chinese in others. Pick one for the batch, or order the second as its own field.
  • A romanized delivery with the scheme left open. Jyutping and Yale disagree on tone marking and on vowels, so an unnamed scheme produces output nothing downstream can rely on.
  • Machine-only transcription with no human pass.

Need Cantonese transcription scoped to your project? Send the audio conditions, volume, and turnaround you need. We will reply with a scoped quote and, if it helps, a free sample.