Cantonese Speech-to-Text in 2026: A Practical Comparison of the Major Tools

If you have ever recorded a meeting in Hong Kong and run it through a transcription app, you know that Cantonese is one of the harder languages to get right. A sentence like "我想去 Causeway Bay 開會 prepare 一下 個 deck" mixes Cantonese and English in a single breath — and that mix is where most transcription tools show their limits.

This is not a niche issue. Code-switching — moving between Cantonese and English mid-sentence — is the everyday register of business, tech, and daily life in Hong Kong. It is also one of the most technically demanding problems in Chinese speech recognition.

We put together a neutral, practical comparison of the major Cantonese speech-to-text tools in 2026 — Google, Apple, Microsoft Azure, OpenAI Whisper, Otter.ai, iFLYTEK, ElevenLabs, and Speechmatics — looking at accuracy, code-switching, privacy, and how easy each one is for a regular person to actually use. Every tool here is good at something; the goal is to help you match the right one to your needs.

How Accuracy Is Measured

Transcription accuracy is usually reported as Word Error Rate (WER) or Character Error Rate (CER) — the percentage of words or characters that differ from a correct reference. Lower is better. Under roughly 5% is excellent; around 20% generally means meaningful manual cleanup.

The most widely cited public benchmark for Cantonese is the Common Voice Yue dataset — real Cantonese speakers reading aloud. On that clean, read-aloud Cantonese, independent testing has reported the following character error rates:

By this benchmark, Apple's on-device Cantonese model scores well — its reported CER is roughly 35% lower than Google's and 67% lower than Microsoft's on this particular dataset, and it runs entirely on-device. For one person dictating clean Cantonese into a Mac, it is a genuinely good, free option.

One important caveat: this benchmark measures clean, read-aloud, single-speaker Cantonese with no English in it. That is not how most people in Hong Kong actually speak.

Why Code-Switching Is the Real Test

Common Voice is careful, monolingual speech. Everyday Hong Kong Cantonese is conversational, fast, and mixed with English:

"個 user feedback 話 個 flow 唔 make sense, 不如 我哋 sync 一 sync 先。"

A benchmark score on clean Cantonese tells you little about how a tool handles a sentence like that. Academic research on Cantonese-English code-mixing points to two reasons this is hard for any engine:

1. English spoken with a Cantonese accent sounds different from "standard" English, which can trip up the English-language model.
2. Limited code-switched training data — many systems were trained mostly on monolingual Cantonese or monolingual English, not the natural mixture.

This affects the entire field, not any single brand. In broader testing, code-switched audio has reduced accuracy across popular tools — one widely reported study saw accuracy drop on Spanish-English mixing. The practical takeaway: for Hong Kong audio, how a tool handles English-in-Cantonese matters more than its headline benchmark score.

The Tools, One by One

Otter.ai

Otter is one of the most popular transcription apps overall, with excellent meeting features and summaries. As of 2026, its transcription is English-focused and it does not offer Cantonese transcription. If you are searching specifically for Cantonese, you will want a tool with native support.

OpenAI Whisper

Whisper supports a very wide range of languages and can transcribe Cantonese. By default it tends to output written standard Chinese (for example "他們" rather than the colloquial "佢哋"). Community fine-tuned Cantonese models exist that preserve colloquial output, but using them involves some technical setup, so Whisper is best suited to developers and power users rather than someone who just wants to open an app.

Google Cloud Speech-to-Text

Google's "yue-Hant-HK" model scales well and is a solid choice for quick drafts and large-volume automated processing within Google Cloud. As with most automated engines, the raw output usually benefits from a review pass, and code-switched English may need cleanup. It is an API/developer product rather than a consumer app.

Microsoft Azure AI Speech

Azure offers strong developer tooling and clean integration with Microsoft cloud workflows, which makes it a popular choice for engineering teams. Its documentation notes that Cantonese output benefits from review where code-switching is involved. Like Google, it is aimed at developers building their own applications.

iFLYTEK (科大訊飛)

iFLYTEK is a leading speech-AI company and has invested heavily in Hong Kong. It now offers Cantonese support directly aimed at the local market — including the iFLYREC Multilingual Conference System (Hong Kong & Macao Edition) and meeting solutions that let users switch among Cantonese, Putonghua, and English, plus on-premises options and translator hardware. It is a capable, serious option, particularly for enterprise and conference deployments where on-premises hosting or dedicated hardware fits the workflow.

ElevenLabs Scribe & Speechmatics

Both are strong, modern speech-to-text engines. ElevenLabs places Cantonese in its higher-accuracy tier, and Speechmatics advertises high accuracy with low latency. Both are primarily cloud, paid, API-first services — excellent building blocks for developers, with audio processed in the cloud.

Side-by-Side Summary

There is no single "best" — it depends on what you are doing. A developer building a product will love Whisper or Azure. A conference organizer may be best served by iFLYTEK's dedicated system. A Mac user dictating clean Cantonese has Apple built in for free.

Three Questions to Pick the Right One

1. How much English is in your audio? If your recordings are mostly clean Cantonese, the headline benchmarks are a fair guide. If they are full of mid-sentence English — like most Hong Kong meetings — prioritize a tool that is specifically built or tuned for code-switching.

2. Where does your audio get processed? Cloud transcription means uploading your recordings to a third-party server. For sensitive meetings, medical, or legal audio, on-device processing keeps the data on your phone or computer.

3. How technical are you? Whisper, Azure, and the API engines are developer tools. iFLYTEK's conference systems are enterprise deployments. If you just want to open an app, press record, and read a transcript, you want a consumer app designed for that.

Where Traverba Fits

Traverba is built for that last group — everyday users who want something simple. No API keys, no cloud account, no enterprise rollout. You open the app, and it works.

The focus is the language people in Hong Kong actually speak: Cantonese with English mixed in. Rather than treating the English as something to strip out, Traverba is tuned on local speech patterns — the mid-sentence English, brand names, and slang of real conversations. And because it processes on-device, your audio stays with you rather than being uploaded to a server.

Traverba is not trying to replace an enterprise conference system or a professional human transcriptionist — those have their place. It is built for the everyday moments where Cantonese and English blur together and you just need readable text, fast, on your own phone.

Try It

Traverba is free on both platforms — voice translation, camera translation, screen translation, and text translation are all free and unlimited.

Available on Google Play and the App Store. Learn more at traverba.com.

Every tool in this comparison is good at something. If what you need is a simple, on-device app tuned for everyday Hong Kong Cantonese-and-English, Traverba is built for exactly that.

Accuracy figures referenced above come from public Cantonese ASR benchmarks on the Common Voice Yue dataset and published research on Cantonese-English code-switching. Feature and language support reflect each provider's offerings as of June 2026 and may change. Real-world accuracy varies with audio quality, accents, and the amount of English in the conversation; we encourage readers to test tools on their own audio.