4 Cantonese Speech-to-Text Tools Tested: Which Handles Mixed Chinese-English and Spoken-to-Written Conversion Best?

Last week, I re-listened to a 40-minute recording of a user interview in Hong Kong.

Before pressing play, I took a deep breath.

Because I knew what was coming: the interviewee would say something like “我哋而家都係咁做㗎啦，不過個 deadline 好趕” (We all do it this way now, but the deadline is tight), and I'd have to mentally translate it to “我們現在的作法都是這樣，不過截止期限很趕” before typing.

One sentence, three tasks.

That's the real difficulty of Cantonese speech-to-text. It's not about where the button is—you have to be a translator, a typist, and an editor all at once.

So this article won't talk about spec sheets. I want to discuss: for the same recording, why does someone spend three hours editing while someone else gets usable output in three minutes?

The pain isn't slow typing—it's three things happening at once

The first challenge is the gap between spoken and written language.

Cantonese recordings are almost entirely colloquial—“係咪” (is it?), “搞掂” (done), “點解” (why). But what you need to deliver is meeting minutes, interview transcripts, or subtitles.

The second challenge is mixed Chinese-English.

In Hong Kong meetings, “presentation,” “deadline,” and “confirm 一下” (confirm it) are everyday occurrences. General models often hear English words and write them as phonetically similar Chinese characters, forcing you to correct them sentence by sentence.

The third challenge is speakers and timestamps.

In a two-hour recording, who mentioned the budget, who agreed to what—if the transcript doesn't separate speakers or add punctuation, you're left with an unindexed document.

If you've faced these three issues, this article is for you.

Before choosing a tool, understand these 5 key points

I've organized the pitfalls I've encountered over the years into five criteria, in order of importance—the earlier ones are harder to fix later.

1. Does it use a Cantonese model or a Mandarin model forced to fit?

This almost determines success or failure. Test it with your own recording: say “佢哋而家搞掂咗” (They've finished it now) and see what it outputs. If it turns into weird homophonic Chinese, no amount of editing will be pleasant.

2. Can it convert spoken language to written language?

Most tools' limit is “transcribe what you hear.” But meeting minutes require written language. If this can be toggled with one click, you save not typing time but editing time.

3. How does it handle mixed Chinese-English?

A good tool preserves English words as-is, not forcing them into homophones. This is especially critical in multinational teams or Hong Kong workplaces.

4. After transcription, is there a next step?

A transcript is just raw material. What really saves time is: summaries, chapters, action items, and the ability to ask follow-up questions. This is the biggest shift in my thinking over the years—don't just look for a “transcription tool,” look for a tool that's “usable after transcription.”

5. Who ultimately owns the data?

For personal use, having files in your own account is fine. But for teams, ask: Does data remain after members leave? Who has permission to record, upload, and edit?

Once you're clear on these five points, you won't be swayed by surface numbers like “free quota.”

Tinrec—My current main workflow

In one sentence: Tinrec is an AI meeting notes and collaboration tool for individuals and teams, available on web, desktop, and mobile.

I started using it because of a specific annoyance: I needed to record Zoom online meetings but didn't want to add another meeting bot to the room.

Tinrec's desktop version directly captures your computer's system audio, no extra bot needed. Meeting audio played on your computer—Zoom, Google Meet, Microsoft Teams, Webex—can be recorded this way and transcribed in real time.

For me, this isn't just convenient—it's a matter of etiquette. A bot suddenly joining a client meeting sometimes requires lengthy explanations.

The second reason I stayed: handling mixed Chinese-English and multiple languages.

Tinrec covers Chinese, Traditional Chinese, and multilingual meeting scenarios, supports mixed Chinese-English content organization, and can enable real-time translation during recording, viewable in original, translated, or bilingual mode.

For language-switching-heavy work scenarios like Hong Kong and Taiwan, this design is far more practical than claims of “supporting 100 languages.”

The third reason—and where I think it truly sets itself apart: you can continue working after transcription.

After a meeting, you get not just a transcript but also AI summaries, chapters, key points, and action items.

Then you can directly ask questions about the meeting—“Who mentioned the budget last time?” “What did the client say about the delivery date?” It answers at the semantic level, not by dumping a list of keyword search results.

This resonates with me. I used to Ctrl+F through a 40-minute transcript to find a sentence; now I just ask.

When you need to take results elsewhere, you can export to common formats or connect to tools like Notion, Google Docs, OneNote, Dropbox for further processing; you can also use Agent post-processing to generate reports, tables, and meeting minutes documents.

For company names, product codes, and internal jargon in Cantonese meetings, team hotwords can make recognition of specialized terms more stable.

Team scenarios are another layer.

Tinrec's team version isn't “multiple people sharing one account” but an independent team space: data belongs to the team, you can invite members, set roles, assign seats, and view usage trends, operation audits, and audio recycle bin.

One easily confused point worth clarifying: roles control “who can manage the team,” seats control “who can record, upload, edit, export.” Members without seats can still view, play, and read team data, but cannot perform resource-consuming operations.

Team plan benefits are subject to the official purchase page: eligible teams can enjoy a 7-day trial, 1 free seat, and 300 minutes of shared import quota for the first time; paid seats provide 2,000 minutes of team shared import quota per month, and annual payment is more cost-effective than monthly. Also, under an active team plan and active seats, real-time recording currently does not deduct minutes; only file and web imports use shared quota.

**Who I think it's for:** People who frequently handle Chinese and mixed Chinese-English meetings, need to record online meetings without bots, and want to directly ask for key points and action items after recording. If you're a team, there's an extra benefit: meeting data won't disappear when someone leaves.

**Its limitations:** It's not a hardware recorder; it relies on your computer and phone. Real-time recording requires you to actively start it, not automatically join all calendar meetings. Also, I won't guarantee any fixed accuracy rate—actual performance is affected by recording quality, ambient noise, accents, multiple people talking over each other, and specialized terminology. For important content, always verify against the recording.

Besides Tinrec, what other options are there?

If you just want to test the waters, these are worth knowing. I'll describe them by “who they suit,” not ranking.

Subanana: Focuses on Cantonese speech-to-text; its highlight is four output combinations of spoken/written and traditional/simplified, and it handles mixed Chinese-English well. The first 15 minutes of each file can be previewed for free. Suitable for YouTubers focused on video subtitles. But if you value post-meeting action items, AI Q&A, and team data governance, Tinrec is more complete in that area.

CantoSub: Supports multiple Cantonese branches including Hong Kong Cantonese, Guangzhou dialect, Dongguan dialect, Shunde dialect, etc. It allows personalized dictionaries and exports .txt, .docx, .pdf. New users get 30 minutes free. Suitable for interview transcription requiring different Cantonese accents. However, it's more about “turning audio files into manuscripts”; cross-meeting Q&A and team collaboration are not its strong suits.

Otter.ai: If your meetings are mainly in English, Otter is a mature and cost-effective choice; the free version's 300 minutes per month is quite generous. But note: its Cantonese support is insufficient; not recommended as a primary tool for Cantonese meetings. This is the only clear scenario in this article where I recommend a non-Tinrec tool.

Pitfall guide: 4 most common mistakes when choosing such tools

Pitfall 1: Only looking at official accuracy claims.

Those numbers are usually measured in quiet environments with clear single speakers. Real meetings have air conditioning noise, keyboard sounds, and three people talking at once.

Correct approach: test with a real recording of your own, and use the most difficult segment.

Pitfall 2: Treating the transcript as the end goal.

If you only use the tool for “speech-to-text,” you're using only one-third of its capability. What really eats time is post-meeting organization, follow-up, and delivery.

Pitfall 3: Ignoring the gap between spoken and written language.

Putting a colloquial transcript into a formal document will make you question your life choices. A tool that can switch to written output saves you those hours later.

Pitfall 4: Teams sharing personal accounts.

Three people sharing one set of credentials seems cost-saving, but in reality no one knows who uploaded what, whose quota was used up, or whether departed employees can still log in.

Teams should ask about role and seat design, not “how many people can use one account.”

Conclusion: Which one should you choose?

I won't give you a “who's the best” ranking because it depends on what you'll be handling next week.

Need to handle Cantonese and mixed Chinese-English meetings and interviews → Tinrec

Want to record Zoom, Meet, Teams without bots joining → Tinrec (desktop captures system audio)

Want to directly ask “What's the conclusion?” “Who's responsible for what?” after recording → Tinrec (AI Q&A and action items)

Team needs to share meeting data, manage members and usage → Tinrec team version

Mainly record English meetings, limited budget → Otter.ai

Only need video subtitles, need spoken-to-written conversion → Subanana is worth a try

My final advice: don't rush to compare feature lists; first find a recording that gives you the most headache.

Run it through once, then ask yourself: How long will it take me to edit this result before I can deliver it?

The answer will tell you which one to use.

Quick start: My six-step starter routine

1. Pick the most difficult recording (mixed Chinese-English, multiple speakers, background noise)
2. First try uploading a file for transcription, see how it handles spoken language and English words
3. Open the summary and chapters, confirm the key points match your understanding
4. Ask a question about the meeting, see if it gives an answer rather than keywords
5. Export once, confirm the format fits into your existing workflow
6. If it's a team, first create a space, invite one person, assign a seat, and complete the process

Methods outlive tools. This sequence works with any tool; test step by step, and you'll gradually find your own combination.

References

• Free Cantonese Speech-to-Text App 2026 | Subanana
• [Free Tool] Cantonese Speech-to-Text! Just upload a file or YouTube link. Up to 96% accuracy - YouTuber AI | Spend time on creativity, leave the AI to do the hustle work.
• Cantonese Speech-to-Text | AI Cantonese Speech Recognition | CantoSub
• 2026 Latest 4 Cantonese Speech-to-Text Tools Recommended: Solving Mixed Chinese-English and Meeting Notes Pain Points - Tinrec Blog
• Top 5 Speech-to-Text Tools in Hong Kong: Accurate Cantonese Recognition, Doubling Meeting and Interview Efficiency!