How to Transcribe Cantonese Audio Accurately

You have a Cantonese recording and need the words on a page: an interview with a grandparent, a radio segment, a sermon, a meeting held in Cantonese. This guide covers the one setting that matters most before you upload, what the transcript looks like when it comes back, and where to spend your proofreading time. Cantonese is harder for speech recognition than Mandarin, and the gap shows in the output, so it also sets honest expectations.

Why Cantonese is harder than Mandarin

Cantonese and Mandarin share a character set, but they are different languages when spoken, and that difference is behind most transcription problems.

Start with what people write. Standard Written Chinese, the form used in newspapers, textbooks and most of the text a speech model learns from, follows Mandarin grammar and vocabulary. A Cantonese speaker in Hong Kong reads and writes it fluently, but it is not what they say out loud. Spoken Cantonese has its own function words and everyday vocabulary: "is" is 係 rather than 是, "not" is 唔 rather than 不, "he" or "she" is 佢 rather than 他, and the possessive 的 becomes 嘅. Verb endings such as 咗 and 緊 mark completed and ongoing actions. Written down, this is called written Cantonese, and it uses characters like 嘅, 咗, 喺, 啲 and 冇 that hardly ever appear in Standard Written Chinese.

A recognition model has seen far more Standard Written Chinese than written Cantonese, so it leans toward the standard forms even when it hears Cantonese sounds. The result can read as if an interpreter had paraphrased the speaker into Mandarin.

Then there are the tones. Cantonese has six against four in Mandarin, and many syllables are told apart by tone alone. A missed tone turns one word into another, and the model's guess will be a plausible-looking character that nobody said.

Finally, Cantonese speakers in Hong Kong, Vancouver, Sydney and London drop English into sentences all the time: a job title, a product name, sometimes a whole clause. Switching languages mid-sentence is harder for any model than one language at a time.

Pick Cantonese explicitly

RocketScribe can auto-detect the spoken language, and for most languages that works well. For Cantonese it is a poor choice. The detector may decide the file is "Chinese", which in practice means Mandarin, and the model then does its best to produce Standard Written Chinese from Cantonese sounds. The output looks like Chinese and may even be readable, but the wording differs from what was said. Particles vanish, 係 becomes 是, and whole phrases are replaced by Mandarin equivalents or by characters that happen to sound similar.

So before you press transcribe, open the language list in the upload box and choose Cantonese. It is a separate entry from Chinese (Mandarin). The Cantonese audio to text page preselects it for you, which is the easiest way to avoid the mistake. If a recording is mostly Mandarin with a few Cantonese sentences, pick Chinese instead and expect to fix the Cantonese parts by hand. The model follows whichever language you set for the whole file.

If you run Whisper on your own computer instead, the same advice applies: set the language to Cantonese rather than letting it guess.

What the transcript looks like

The transcript comes back in Chinese characters, as paragraphs with timestamps. With Cantonese selected, most of the colloquial words come through as written Cantonese, so you will see 係, 唔, 佢 and 嘅 where they belong. It is not perfectly consistent. In some lines the model falls back to the standard form, writing 是 where the speaker said 係, or drops a sentence-final particle such as 呀 or 啦. If your project needs the exact wording, for example a quoted interview or linguistic fieldwork, treat those as items to check rather than assume the model got them right.

The script can vary as well. Most Cantonese output is in traditional characters, matching Hong Kong usage, but simplified characters can appear, and RocketScribe does not convert between the two. If you need one script throughout, run the exported text through a converter.

English words usually appear in Latin letters. Now and then one is misheard as a Cantonese syllable and written as a character, or a Cantonese word comes out as English. Both are quick to spot once you listen to the line.

Fixing names, slang and English in the editor

Names are where the model guesses hardest, because it cannot know which of several same-sounding characters a family uses. Play the line, type the correct name once, then use your browser's find function to locate the wrong version elsewhere.

The player under the text highlights the line being spoken and jumps to any line you click, so the quickest way to proofread is to play the recording and read along, pausing where the text and the sound disagree. Click into a line to edit it in place, and split or merge paragraphs where the model broke them in the wrong place.

Slang deserves a second look too. Expressions that spread through forums and chat groups rarely appear in the text the model learned from, so they tend to come out as characters with the right sounds and the wrong meaning. When the text reads the way the speaker talks, export it as TXT or DOCX for notes and quotes, or as SRT if the recording is a video you want to caption.

Recording tips that matter most for Cantonese

Cantonese accuracy depends more on clean audio than Mandarin does, so the recording itself is where you gain the most.

• Keep the speaker within arm's length of the microphone. A phone on a table between two people picks up the room more than the voices.
• Turn off the television or radio. A Cantonese drama in the background is a second Cantonese speaker as far as the model is concerned.
• For family interviews, ask one question at a time and let the answer finish. Overlapping voices are the fastest way to lose accuracy.
• Upload the original file. A voice note re-shared across several apps has been re-encoded each time, and each pass loses a little more.
• Video is fine as it is; upload the MP4 or MOV rather than converting it first.

Honest expectations

Fast conversational speech between several people, heavy slang and a speakerphone in an echoing room each lower accuracy, and together they can produce a transcript that needs line-by-line correction. A single clear speaker close to the mic gives a draft that mostly needs names and particles fixed. Proofread either way, with the audio playing.

Speaker identification (the Identify speakers switch, on paid plans) groups lines by voice and labels them Speaker 1, Speaker 2 and so on. It does not know who anyone is, so the names come from you; click a label to rename it everywhere at once.

Try it on your own recording

Upload a file on the Cantonese audio to text page, where Cantonese is already selected, and read the result with the player running. If you remember one thing from this guide, pick the language yourself.