Chinese transcription is a procurement question before it is a file upload. Ask which spoken variety the model handles, which script it writes, and whether the result preserves speech or normalizes it into formal written Chinese. A single “Chinese” setting may leave those decisions implicit.

Key Takeaways

Chinese transcription has three separate requirements

Spoken variety tells the recognition system what speech to expect. Mandarin and Cantonese need separate evaluation; a model that accepts a generic Chinese code may not provide the same results for both. Script asks whether output should use Simplified or Traditional characters. Register asks whether a Cantonese transcript should retain vernacular wording or read like Standard Written Chinese. These are distinct deliverable requirements, even when a service combines them in one setting.

A formal rewrite may suit meeting notes, but it cannot serve as a verbatim quotation without checking the recording. For an interview, legal record, or research transcript, ask for a sample that preserves the wording you need and keep the source audio available for review.

Need a wider tool comparison? Our multilingual transcription guide compares live translation with post-meeting transcripts. This page focuses on the Chinese language choices inside the transcription job.

What a Cantonese benchmark label can hide

An ACL 2025 dataset audit classified the written prompts attached to Cantonese-labelled speech subsets. Its Table 5 found a sharp difference between Common Voice 17.0 yue and FLEURS yue_hk:

Text promptsStandard Written ChineseWritten vernacular CantoneseUnmarked
Common Voice 17.0 yue0.1%75.7%24.1%
FLEURS yue_hk89.8%0%10.2%

These percentages classify text prompts, not the spoken variety of individual recordings. Common Voice also has a small mixed category omitted from this table.

The paper suggests that FLEURS yue_hk may be mislabeled because its prompts largely use Standard Written Chinese. That does not establish that no speaker pronounced the recordings in Cantonese. It means the label alone is insufficient evidence for a claim about colloquial Cantonese transcription. Ask whether a vendor’s benchmark uses read prompts, conversational speech, or your target domain.

A separate LREC 2022 Cantonese dataset study found 10.15% character error rate when a model trained on its MDCC corpus was tested on MDCC, but 83.42% when the same model was tested on Common Voice zh-HK. This is a result for one architecture and two test sets, not a universal Cantonese error rate. The practical lesson is to test your accents, microphones and speaking style.

A 2026 multi-dialect ASR study likewise reports different results on Mandarin and its dialect test sets, but those numbers cannot be compared directly with the 2022 Cantonese experiment. For more on error metrics, see our AI transcription accuracy comparison.

Read language codes as model-specific promises

Google Cloud Speech-to-Text lists codes including cmn-Hans-CN for Simplified Mandarin, cmn-Hant-TW for Traditional Mandarin and yue-Hant-HK for Traditional Cantonese. These codes bundle variety, script and region. Which recognition models support each code also varies by location.

Deepgram’s model list labels Cantonese Traditional as zh-HK for Nova-3, alongside Simplified and Traditional Mandarin options. Its Flux STT language list does not include Chinese. A language shown for one model should never be assumed to work in another.

Neither code list tells you whether colloquial Cantonese will be normalized into formal prose. Request a short sample in the intended script and register. If the supplier cannot answer the register question, compare the sample line by line with the audio.

Compare the service, not just the headline rate

As of October 2026, ElevenLabs’ Chinese Scribe page lists $0.40 per audio hour, while GoTranscript’s Chinese page advertises human transcription from $1.02 per minute. These are starting prices for different scopes, not equivalent quotes. The human service advertises Mandarin and Cantonese handling, Simplified or Traditional deliverables, and optional localization. Verify current rates, add-ons and the exact deliverable before ordering.

Benchmark numbers need similar care. AssemblyAI’s vendor-run multilingual comparison presents one Chinese column rather than separate Mandarin and Cantonese rows. It is a snapshot of the tested systems and datasets, not proof of performance on your Cantonese interview. Ask which model version, language subset and character error rate calculation produced a quoted score.

For Mandarin-English switching, build a bilingual sample too. The ISCSLP 2022 Chinese-English challenge reported 16.70% mixture error rate for its winning system versus a 26.50% challenge baseline. That controlled result cannot predict every meeting, but it shows why a monolingual demo is an incomplete purchasing test.

A four-step Chinese audio-to-text test

  1. Identify the speech. Record whether the source is Mandarin, Cantonese, another variety, or mixed with English. Select an explicit supported code where the model offers one.
  2. Specify the output. State Simplified or Traditional characters and whether the transcript must preserve conversational wording.
  3. Test representative audio. Include real accents, background noise, names and any language switching. Inspect the output alongside the recording.
  4. Keep a review path. Save the audio, document the model and settings, and budget human review when exact wording is material.

Live captions and an archival transcript have different success criteria. See real-time versus recorded-audio transcription for that workflow choice. If the goal is Chinese text translated from another spoken language, use a speech translation workflow such as our Arabic to Chinese translator guide.

Where MirrorCaption fits

MirrorCaption currently shows one Chinese language option. It does not offer a separate Cantonese option or a Simplified/Traditional script control, and we publish no Chinese character error rate for the product. Uploaded recordings take one selected source language per file. Use a service with explicit controls and human review when your job requires a certified or verbatim Cantonese record.

For live Mandarin-English conversation, MirrorCaption can show the original speech beside a translation while people are talking. Try it with your own sample to decide whether that is useful for the meeting; it does not replace a reviewed archival transcript.

Try MirrorCaption Free

See live Chinese and English captions side by side on a representative conversation. One free hour to try, one-time, with no credit card.

Get Started Free

Frequently Asked Questions

Is Cantonese transcription as accurate as Mandarin?

There is no single comparable accuracy number. A 2022 Cantonese study showed one model moving from 10.15% character error rate on its own test set to 83.42% on another corpus. Test the same kind of audio you plan to transcribe.

Can transcription software tell Mandarin and Cantonese apart automatically?

Detection depends on the model and recording. Google Cloud lists Cantonese as yue-Hant-HK; Deepgram Nova-3 lists it as zh-HK. Choose an explicit variety when available and test a sample.

Will a Chinese transcript come back in Simplified or Traditional characters?

It depends on the service and model. Some documented language codes bundle the script with a variety and region. Confirm the actual output before submitting a large batch.

How do tools handle a meeting that mixes Chinese and English?

Use a representative recording with language switches and names. The ISCSLP 2022 challenge reported 16.70% mixture error rate for its winning system on its own test set, which is a reminder to test bilingual audio directly.

Does MirrorCaption let me choose the Chinese variety and script?

No. MirrorCaption currently offers one Chinese language option, without separate Cantonese or script controls. We do not publish a Chinese accuracy benchmark of our own.