Dictating mixed Chinese and English on a Mac
Say "这个 bug 明天 fix 一下" and most dictation tools have to guess one language for the whole sentence, so half of it comes out wrong. Here is why, which models get it right, and how to set it up on a Mac.
Short answer: use a speech model that was trained on Mandarin–English speech, and don't force one language. In Vocco that is Qwen3 ASR, the default download. Hold right ⌘, speak the way you normally do, release, and the mixed sentence is pasted at your cursor. It runs offline.
On 300 utterances of real Mandarin–English conversation, Qwen3 ASR made about 16 errors per 100 characters and words, against 21 for Apple's on-device recognizer and 28 for Whisper large-v3-turbo. On English words alone the gap is wider: 29% wrong for Qwen3, 53% for Apple.
Why dictation breaks on mixed sentences
Most dictation tools decide on one language per dictation, then decode everything as that language.
- Wispr Flow: "Flow detects one language per dictation, not per word," and with Chinese and English both selected, "English words can appear in Chinese characters or vice versa" (Wispr help).
- superwhisper recommends setting your language instead of Automatic, and using Automatic "only when you mix languages inside a single dictation" (superwhisper docs).
- macOS Dictation lets you switch languages with the Globe key or the language next to the cursor (Apple support).
- MacWhisper: "not able to transcribe files with multiple languages at the same time" (MacWhisper docs).
When the language is pinned to Chinese, English words get forced into Chinese characters. An early Vocco build had exactly this bug: Whisper, pinned to Chinese, wrote the English word "meeting" as "秘诊". When we measured it, Whisper Base pinned to Chinese got 92% of English words wrong, and on the English-only sentences it made more errors than there were words. Letting it detect the language brought English word errors down to 51%. A bigger model brought it lower still, and a model trained on mixed speech did better again.
Which engines handle it
The first three rows run on the Mac inside Vocco exactly as measured. The Apple row is Apple's newest on-device recognizer, and the Whisper row runs it with automatic language, the way many Mac dictation apps do; both are there for comparison. Lower is better. MER counts errors per Chinese character and per English word together.
| Engine | Overall MER | English words wrong | Mixed sentences MER |
|---|---|---|---|
| Qwen3 ASR (Vocco default) | 0.157 | 29% | 0.170 |
| Paraformer Large (in Vocco) | 0.170 | 35% | 0.145 |
| SenseVoice Small (in Vocco) | 0.185 | 35% | 0.184 |
| Apple on-device recognizer, zh-CN | 0.207 | 53% | 0.207 |
| Whisper large-v3-turbo, auto language | 0.277 | 37% | 0.232 |
Paraformer is slightly better than Qwen3 on the 62 sentences that actually mix languages, and the two are within noise of each other overall. Qwen3 is the default because it is much better on English words and also recognizes Japanese, Korean and French. If you mostly speak Chinese with a few English terms, Paraformer is worth a try; both are in Vocco.
Full method, model builds, per-subset numbers and example failures are on the benchmark page.
Set it up in Vocco
- Download Vocco and open it. There is nothing to sign up for.
- Stay online for a few minutes the first time: Vocco downloads Qwen3 ASR (987 MB) in the background and makes it the engine for dictation, live and file transcription by itself.
- Click into any text field, hold right ⌘, speak, and release. The text is pasted at the cursor and stays on the clipboard.
- For a meeting or a lecture, use the Real-time window. For a recording, drop the file into Files.
There is no language setting to change. Qwen3 ASR detects the language as it goes.
What it looks like
Real utterances from the ASCEND test set, with what two of Vocco's engines wrote. Mistakes are left as they came out.
Limits
- Chinese output is Simplified. There is no Traditional Chinese output yet.
- Our benchmark is Mandarin–English conversation. Cantonese–English and other pairs are not measured.
- Very short fillers like "嗯" or "okay" are the most common errors for every engine.
- Your own project names may not come out right; Vocco has no custom vocabulary yet.
Questions
Does macOS Dictation support Chinese and English in one sentence?
Apple documents switching between dictation languages, by clicking the language next to the cursor or pressing the Globe key. It does not document recognizing both inside one sentence. Add both languages under System Settings → Keyboard → Dictation if you want to switch quickly.
Why does my dictation app write English words in Chinese characters?
Because it is decoding the whole utterance as Chinese. Wispr Flow's help page says the same thing happens there when Chinese and English are both selected. A model trained on code-switched speech, or one that detects language within the sentence, avoids it.
Does Vocco need an internet connection for this?
No. Qwen3 ASR runs on the Mac. You need a connection once, to download the 987 MB model, and no account.
Does it work for Cantonese and English?
We have not measured Cantonese–English mixing, so we don't claim it. Our benchmark is Mandarin–English.
How this page was made: numbers and examples come from our benchmark run of October 22, 2026; competitor behaviour is quoted from their own documentation on October 22, 2026. Drafted with AI assistance and checked against those sources.
Try mixing languages on your Mac
Download, hold right ⌘ and say a sentence that switches language halfway. Release and read it back. Free, no sign-up.
macOS 14 or later · Apple silicon