Cantonese speech to text with Potok comes with a borderline note. Published tests of large-v3 report a character error rate of 9.8, since Cantonese is written without spaces, but a vendor test showed much weaker results for the faster turbo variant. Treat the numbers as other people's findings, not ours.
Transcribe a recording
Send your voice message, audio or video to @potok_now_bot on Telegram and the text comes back in the chat.
Open @potok_now_bot in Telegram
Files up to 2 GB · export to TXT, MD, DOCX, PDF, SRT and VTT · speaker labels and timestamps · about 30 languages
Key facts
- Quality is quoted per character, not per word, because Cantonese text has no spaces between words.
- A vendor test reported much worse results for the turbo variant, which is why the language is flagged as borderline.
- Edge status means the figure is close to our cut-off, so a sample test is worth the minute it takes.
- Exports stay in the language of the recording.
What a transcript looks like
[00:00] Speaker 1: 我哋開始開會啦。首先睇下上個禮拜嘅結果，跟住傾下一步點做。Meaning in English: Let's start the meeting. First, we'll review last week's results, then plan the next steps.
Written by us to show the script and the layout of a transcript. It is not the output of the recognition engine.
Published quality
| Language | Verdict | Published result | Notes |
|---|---|---|---|
| Cantonese (粵語) | Recommended, borderline | character error rate 9.8% (large-v3, FLEURS) Whale paper | Published tests show an error rate close to our 15% line. Check a short sample first. Cantonese is written without spaces, so quality is reported as a character error rate. A vendor test reported much worse results for the faster turbo variant, so treat this language as borderline. |
These numbers come from published tests by other people, not from tests by Potok. Your own audio, with its noise, accents and microphone, will differ.
How to do it
- Start with a test clip. Send thirty seconds first. For a borderline language this costs little and shows whether the output suits you.
- Use the bot. Open @potok_now_bot and attach the sample as a voice message or audio file.
- Judge by character. Read the text beside the audio and count the places where a character is wrong, not whole sentences.
- Send the rest. If the sample reads well, send the full recording. Up to 2 GB is fine and the Upload button handles big files.
- Choose an export. DOCX or TXT for reading, SRT or VTT for captions.
| Feature | Details |
|---|---|
| Where it works | Telegram bot @potok_now_bot; large files go through the bot's mini app |
| File size | Up to 2 GB |
| Export | TXT, MD, DOCX, PDF, SRT, VTT |
| Speakers and time | Speaker labels and timestamps can be switched on |
| Languages | About 30 languages with good published results; a few more with warnings |
| Translation | The transcript is not translated into another language |
| Dictation | A separate Potok desktop app lets you dictate text into any program; it recognises Russian speech |
Common mistakes
| Mistake | What happens | Fix |
|---|---|---|
| Reading a character error rate as a word error rate | 9.8 counts characters, so the two are not comparable. | Compare only figures that use the same metric. |
| Committing a long file without a test | A borderline language can surprise you on a whole hour. | Always run a short sample first. |
| Mixing English phrases into the same clip | Foreign words are written less reliably. | Mark those spots and check them by ear. |
| Expecting spaces between words | Written Cantonese does not use them. | Do not search the transcript for word breaks. |
More
See the full list of languages, the tools and the glossary.
Frequently asked questions
Why is Cantonese borderline?
A vendor test reported much worse results for the faster turbo variant, so the language sits near our cut-off.
What is a character error rate?
It is the share of characters that differ from a reference text, used when writing has no spaces between words.
Is 9.8 our own measurement?
No. It comes from published tests of the large-v3 model, run by others.
Can the bot handle a video?
Yes. Video, audio and voice messages are all accepted, up to 2 GB per file.
Will speakers be named?
No. Voices can be labelled with numbers only, never with real names.
Sources
Each source was opened by hand on the date shown. Pages change, so open the source if a detail matters to you.
- Word error rate (Wikipedia), Wikipedia · checked
- Whisper speech recognition system (Wikipedia), Wikipedia · checked
- Whale paper, Table 5 (Whisper large-v3 on FLEURS) · checked
Try it on your own recording
Send your voice message, audio or video to @potok_now_bot on Telegram and the text comes back in the chat.
Open @potok_now_bot in Telegram