Why Cantonese audio to text needs its own setting
Cantonese and Mandarin share a writing system but sound very different. If you leave a Cantonese file on auto-detect, the speech model may decide it is Mandarin, and you get a transcript that reads like a rough paraphrase into standard written Chinese with many wrong words. That is why this page preselects Cantonese in the upload box.
You can still change the language before each upload. If a recording is mostly Mandarin with some Cantonese, pick the language that covers most of the speech and expect to fix the other parts by hand.
What to expect from the transcript
Cantonese is supported, but the model has seen much less Cantonese than Mandarin during training. Results depend more on clear audio than they do for Mandarin, and you should plan to read the whole transcript with the recording playing. The synced player highlights the current line and lets you click any line to jump there, which makes this review quicker.
The output is in Chinese characters. Spoken Cantonese has many words and particles that people write in their own way, and the model does not always write them the way a Cantonese speaker would type a message. Sometimes it uses the standard written Chinese form instead of the colloquial character. Sentence-final particles may be dropped or changed. If your project needs the colloquial wording exactly as spoken, for example for linguistic research or a quoted interview, budget time to correct these.
RocketScribe does not convert between Simplified and Traditional characters. If the transcript comes out in a different script from the one you need, use a conversion tool on the exported text.
Recordings that work better
Because Cantonese results are more sensitive to audio, small changes before recording help:
- Put the phone or recorder close to the person speaking, ideally within arm's length.
- Avoid rooms with a TV, radio or restaurant noise in the background.
- For interviews with older relatives, ask one question at a time so voices do not overlap.
- If you already have the recording, upload the original file rather than a copy that has been re-shared through several chat apps, since each step can lower the quality.
Mixed Cantonese and English
Cantonese speakers in Hong Kong, the UK, Canada and Australia often drop English words into a sentence: a job title, a brand, an email, a deadline. These words often come out in Latin letters, but switching languages mid-sentence is harder for the model than one language at a time. Expect some English words to be misheard or rendered in Chinese characters, and check them during review.
The same applies to Mandarin. Many Hong Kong and diaspora speakers move between Cantonese, Mandarin and English in one conversation. Pick the language spoken most and correct the rest by hand, or split a long recording into parts if whole sections are in a different language and run each part with the right setting.
What people do with Cantonese transcripts
Common uses include family history recordings, oral history projects about Hong Kong and Guangdong migration, academic interviews, and subtitles for community or church videos. Export TXT or DOCX for reading and archiving, SRT or VTT for subtitles, or CSV to work through the transcript in a spreadsheet. These formats keep every character. The direct PDF supports Western European characters only, so for Chinese text the PDF option opens a print-ready page you save from your browser.
Relatives or colleagues who do not read Chinese can use the Google Translate link to follow along in English. The AI prompt export copies the transcript with a ready prompt for ChatGPT or Claude, useful for a quick summary of a long recording. Speaker identification is available on paid plans or with a Minute Pack.