Chinese meeting transcription: Mandarin, Cantonese, and which script you get back

The short answer

1. Noter AI wins on coverage. Chinese and Cantonese both sit inside its 60+ transcription languages, automatic language identification runs on every recording, and the finished notes can be produced in Simplified or Traditional characters. It also records a meeting held in a room, which most of this list cannot.

2. Fireflies.ai has the widest catalogue at 100+ languages with auto-detection, and it handles a Mandarin call that keeps sliding into English. It records in person from its mobile app too. The AI summaries are rationed by monthly credits, and pricing is per seat.

3. Notta is genuinely strong here. Mandarin and Cantonese are both on its 58-language list and the bilingual side-by-side view suits Chinese-English minutes. You still have to choose the language before you press record.

4. Granola lists Mandarin and Cantonese, which is unusual for a smaller tool. There is no bot: the iPhone and Android apps record meetings in a room, and the Mac desktop app captures system audio for virtual calls. Cantonese is on the desktop language set and not the mobile one.

5. tl;dv lists Chinese (Mandarin) and nothing for Cantonese. Good sales-call analytics, weak iOS app, free plan capped at 10 AI summaries for the life of the account.

6. Fathom has the best free plan in the category and no Chinese on its published transcript language list at all. Thirty-eight languages, Chinese absent.

7. Otter.ai supports Chinese Simplified only, as one of six languages. There is no Cantonese, no automatic detection, and nothing that produces Traditional characters or a translated summary. It gets its own section below because a lot of people assume it covers more than it does.

Why Chinese audio goes wrong in ways you can miss

• Homophones land inside contract vocabulary. 权利 (rights) and 权力 (power) are both quánlì. 定金 and 订金 are both dìngjīn and mean different things in a deposit clause. The model picks characters from context, and legal or commercial terms are exactly where a wrong-but-fluent guess slips past a proofread.
• Cantonese is not accented Mandarin. It has six tones instead of four and its own written characters: 係, 唔, 咗, 嘅, 佢, 哋. A model that only knows Mandarin will render a Cantonese meeting as Mandarin-shaped text that nobody in the room said. Check that Cantonese is listed by name on the language page, since a generic Chinese entry usually means Mandarin.
• Which script you get back decides who can read the notes. 会议纪要 and 會議紀要 are the same document to a machine. A reader in Taipei or Hong Kong sees Simplified minutes as careless work, even when every character in them is technically correct.
• English is welded into the grammar. 呢個 project 我哋下星期 follow up. 这个 deadline 我们下周再 confirm. Any tool that locks to Chinese when you hit record writes those English words out phonetically in characters, which is how a follow-up date ends up unsearchable.
• Numbers and surnames absorb your correction time. 十四 and 四十 flip in fast speech, and 万 as a counting unit turns figures into something else entirely if the model mishears the classifier. Surnames collide constantly too: 张 against 章, 李 against 黎. Record five minutes of your own audio before subscribing to anything on this page, and read the numbers first.

All 7 apps compared

Column two is where the famous names thin out, and column three is where nearly all of them stop. Column five matters more than people expect, because supplier visits, factory walkthroughs and negotiations over dinner are never scheduled video calls, and a bot-based tool cannot hear any of them.

#1 Noter AI

Noter AI is this site's own app, so read this section with that in mind. The reasons it leads for Chinese are specific and checkable.

Cantonese has its own entry on the language list. Transcription runs on Soniox's unified speech model, which lists Chinese and Cantonese as separate entries inside a 60+ language set. A Guangzhou or Hong Kong meeting is not forced through a Mandarin model and then quietly corrected into something plausible.

There is nothing to select before recording. Automatic language identification is enabled on every recording, so you never set Mandarin or Cantonese by hand, and a bilingual client meeting does not have to be predicted.

Code-switching stays intact. Language is tagged at the word level across the whole file, so 呢個 project 下星期 follow up comes back with the Chinese written as Chinese and the English written as English. On a tool that commits to one language at record time, half that sentence turns into invented characters.

It records the room. One tap on an iPhone, iPad or Android phone, a Control Center widget, or Siri, and it keeps running in the background. You can also import an audio file you recorded some other way. For virtual meetings a bot joins Zoom, Microsoft Teams, Google Meet and Webex.

Simplified or Traditional, whichever your readers need. The executive summary, detailed analysis and action items can be produced in another language, and Chinese (Simplified) and Chinese (Traditional) are both among the 18 output languages. A Mandarin call can go to a Taipei team in Traditional characters, and an English board meeting can come back as a Chinese summary for the mainland office.

One flat price. $9.99/month or $49.99/year covers transcription, translation, summaries and action items, with no per-seat fees and no monthly AI credits. Rated 4.8★ on the App Store.

The honest limits. The bot joins Zoom, Teams, Meet and Webex, and it cannot join Tencent Meeting, DingTalk or Feishu. For those, record the speaker audio from the phone sitting next to the laptop, or import the recording afterwards. Homophone errors on contract terms and surnames still happen on every tool here, this one included. Script comes from the translation step, so check the raw transcript on your own audio if your team reads Traditional throughout. There is also no web or desktop app and no CRM integrations. Nor is there a permanent free tier, though roughly five minutes of your own audio can be tested free first.

#2 Fireflies: the biggest language list, metered summaries

Fireflies covers Chinese inside 100+ languages with automatic detection, so it avoids the pick-a-language trap and copes with a Mandarin meeting that keeps borrowing English. With 50+ integrations pushing notes into a CRM, it is the sensible choice for a company whose meetings all live on Zoom or Teams.

What keeps it off the top: its AI summaries are capped by monthly credits even on paid plans, pricing is per seat, and the summary never crosses a language boundary. Its Cantonese position is also unclear, so verify it against your own audio before you count on it. See the Fireflies comparison.

#3 to #6: the rest

#3 Notta is the strongest non-Noter option for Chinese specifically. Mandarin and Cantonese are both covered, transcript translation reaches 42 languages, and the bilingual side-by-side view is the best in the category if your job is producing Chinese-English minutes. The blocker is structural: there is no automatic language detection, so you commit to a language before the meeting shows its hand. Free sessions are capped at three minutes and the AI training default uses your data on non-Enterprise plans. See the Notta comparison.

#4 Granola lists both Mandarin and Cantonese, which is more than most tools its size manage. It works without a bot: the Mac app captures system audio for virtual calls and the mobile apps record a meeting in a room. Cantonese appears on the desktop language set and not the mobile one, which is the wrong way round for a site visit in Kowloon. See the Granola comparison.

#5 tl;dv lists Chinese (Mandarin) and does not list Cantonese. It is built around sales-call coaching and scorecards. Its iOS app sits around 2★ and the free plan allows 10 AI summaries for the lifetime of the account. See the tl;dv comparison.

#6 Fathom has the most generous free plan of anything here, and its published transcript list runs to 38 languages with no Chinese entry. It also records online calls only and cannot import an audio file. Confirm against Fathom's current list before you write it off, but as things stand it is not a Chinese tool. See the Fathom comparison.

#7 Otter.ai supports Mandarin only, with limits

Otter is the name people reach for first, and its Chinese support is narrower than the reputation suggests. It covers six languages: English (US and UK), Spanish, French, German, Japanese, and Chinese Simplified.

So Mandarin does work, within limits. There is no Cantonese and no automatic language detection, so you set the language by hand and a bilingual meeting becomes a coin flip. There is no Traditional-character output either, and Otter never translates the summary, so a Mandarin meeting produces Mandarin notes and nothing a head office elsewhere can read.

Otter's live scrolling English transcription is still the best in this comparison, which is of no use to a Cantonese meeting in Kowloon. See the Otter.ai comparison.

Otter and Fathom sit at the bottom for opposite reasons: Fathom does not list Chinese at all, while Otter lists it and then does very little with it.

Getting a cleaner Chinese transcript

1. 1Test on five minutes of a real meeting, with your own product names and your usual amount of English in it. A clean studio sample tells you nothing about a factory conference room.
2. 2Put the phone flat in the middle of the table. Tone is carried in pitch contour, and pitch is the first thing a muffled or clipped recording loses.
3. 3Confirm which Chinese a tool means. If Cantonese is not named on the language list, assume it is not covered, whatever the total number on the pricing page says.
4. 4Decide the output script before you circulate anything. Set the notes to Traditional for a Taipei or Hong Kong audience and Simplified for the mainland. Converting characters by hand afterwards costs far more time.
5. 5Proofread figures, dates and surnames first. That is where nearly all of your correction time goes, and a wrong 万 or a wrong 章 reads as perfectly ordinary Chinese.
6. 6Do not ask anyone to stop using English words. With auto-detection it is unnecessary, and no Hong Kong or Singapore meeting has ever managed it anyway.

Bottom line

For Chinese, the language list on a pricing page settles almost nothing. The questions that matter are whether Cantonese is named, whether the tool has to be told the language in advance, whether it can record a meeting held in a room, and which script the notes arrive in.

Fathom does not list Chinese. Otter gives you Mandarin, chosen by hand, and no Cantonese. tl;dv covers Mandarin and skips Cantonese. Notta covers both and makes you choose before the meeting starts. Granola covers both and keeps Cantonese on the desktop app only. Noter AI covers Mandarin and Cantonese with detection always on, records in person or on a call, and produces the summary and action items in Simplified or Traditional characters, at $9.99/month or $49.99/year. Test five minutes of your own audio, then read the numbers.