Missing punctuation in Chinese Transcription

[RobZuazua]

Which Deepgram product are you using?

Deepgram API

Details

I notice that punctuation works well when transcribing English.

However, when the input file is Chinese, with detect language set to true, the response missed probably 85% of the necessary periods.

Is this something you all are aware of / is there a fix coming?

If you are making a request to the Deepgram API, what is the full Deepgram URL you are making a request to?

POST /v1/listen?punctuate=true&model=whisper-large&detect_language=true

If you are making a request to the Deepgram API and have a request ID, please paste it below:

42bf5903-6db5-407a-95c9-07d98d60d8b7

If possible, please attach your code or paste it into the text box.

async function submitAsyncRequest(audioSource) {
const response = await deepgram.transcription.preRecorded(audioSource, {
punctuate: true,
model: "whisper-large",
detect_language: true,
// tier: "nova",
// other options are available
});
return response;
}

If possible, please attach an example audio file to reproduce the issue.

No response

[jkroll-deepgram]

Hi @RobZuazua, Deepgram does support punctuation more robustly in English than in Chinese. However, I also recommend trying smart_format=true instead of punctuate=true. Smart Formatting improves readability of the transcript overall, including general punctuation and grammar. You may still notice some gaps in formatting for Chinese, but that will give you the currently best available results. We continue to actively work on improving formatting across the many languages that we support.

[RobZuazua]

appreciate the input

[RobZuazua]

fwiw: smart_format=true did not help

[jkroll-deepgram]

Sorry to hear that smart formatting didn't help 😕 It includes punctuation as well as other text formatting, so smart_format=true does everything that punctuate=true does, plus more (but it varies by language). Given what you're observing, our Chinese formatting is just not on par with English yet.

I also see you're using whisper-large model. You could also try Deepgram's own base model which includes Chinese in a public beta. This would be model=general&tier=base. We support three related language codes: zh (Chinese), zh-CN (mainland China-specific), and zh-TW (Taiwan-specific). It's worth trying out the base model, but you'd have to see how well it works on your data particularly also using language detection.

Thanks for using Deepgram and I hope you can settle on parameters that give you sufficient performance for your use case!