Claude Code Chinese Voice Input on Mac: Support and Setup

Does Claude Code /voice support Chinese?

Chinese is absent from the 20 supported dictation languages in Anthropic’s voice guide, checked October 17, 2026. An unsupported language setting falls back to English for dictation; text responses are unaffected. Chinese word counting is mentioned elsewhere on the page, but is not a transcription-support statement.

Do not confuse Claude Code CLI dictation with the Claude chat app’s display or voice settings. Anthropic documents those language settings separately. A Chinese response from Claude is not evidence that /voice is configured to recognize Chinese.

This is a documented-support check, not a claim that every Chinese utterance will fail. We have not benchmarked native /voice against Ducduc. If your installed Claude Code transcribes Chinese, test your own project terms and consult the current official list rather than relying on an old setup post.

A Mac route for Chinese-English prompts

Ducduc is a separate Mac dictation app. It records on the Mac, recognizes Chinese-English speech on-device, and inserts text into the local Claude Code input. It needs macOS 14 or later, permissions, and the required speech resources; no Claude Code plugin is required.

1. Download Ducduc and complete its Microphone, Speech Recognition, and Accessibility setup. These permissions belong to Ducduc; native /voice requests microphone access for the terminal instead.
2. In Settings → Dictation, choose Recognition language → Chinese (with mixed English). If the app reports missing resources, follow the language-model guide.
3. Check the actual Dictate key in that tab. A fresh installation may choose Right Command when macOS owns fn. Use the shortcut guide if fn opens emoji or changes input sources.
4. Enable Write prompts in AI coding tools when you want the spoken brief organized as an instruction. Ordinary dictation is sufficient; the separate AI Command mode has optional automatic submission.
5. Turn native Claude Code dictation off with /voice off for this check. Focus the Claude Code prompt, start Ducduc, speak, and stop. Keep the same field focused until delivery finishes, then review and send yourself.

Using one recorder for the test makes it clear which app produced the text. You can choose a different input method later.

Check meaning and identifiers in one realistic prompt

Use this synthetic test phrase as a template, replacing the term with one from your project. It is not a captured output or a recognition benchmark:

先检查 login callback 为什么触发两次，我怀疑是 retry，但还不确定。先给我原因和修改方案，不要直接改数据库 schema。

Before sending, inspect four separate requirements:

Check	Why it matters
login callback, retry, and schema survive	The agent must know which technical concepts you meant
“I suspect” stays uncertain	A guess must not become an established cause
Investigate and propose a change first	Diagnosis must precede implementation if that is what you asked
Do not change the database schema	A late spoken constraint still applies to the whole request

Ducduc’s instruction cleanup aims to organize the request and remove repetition while retaining those relationships. A short request can stay a paragraph; a list is useful only if the content calls for one. Check the result instead of judging cleanup by how many bullets it adds.

Speech is less suitable for exact casing, punctuation, or long paths. If LoginCallbackV2 is the actual symbol, type or paste it into the prompt after dictating the surrounding explanation. A personal dictionary can help supply recurring vocabulary; it cannot guarantee the exact spelling of every identifier.

Diagnose the stage that failed

What you see	Next action
No listening indicator	Verify Ducduc’s configured key and Accessibility permission; resolve fn conflicts before changing languages
The indicator reacts, but there is no transcript	Check microphone input and the downloaded recognition language
The final prompt has a wrong English term	Open Dictation History → Show raw transcript and compare the two versions
The transcript is correct, but no text reaches Claude Code	Read the copy card, check focus and Secure Keyboard Entry, then use Copy and paste manually

If the raw transcript already contains the error, start with mixed Chinese-English recognition and personal vocabulary. If only the cleaned text is wrong, changing the microphone will not fix that transformation: review cleanup settings and keep a short non-sensitive raw/final pair for feedback.

What this workaround does and does not keep local

Ducduc’s recognition audio stays on the Mac. Default built-in cleanup sends recognized text to Ducduc’s service; a remote BYOK provider also receives text. For local cleanup, use a downloaded local Ollama model, disable Ollama cloud features and configure Ducduc Offline Mode. A localhost endpoint alone does not establish local inference. The local setup guide and data-flow policy explain the boundaries.

For SSH, Ducduc supplies text to the local terminal client; it does not forward your microphone to the remote host. Whether the client accepts the paste depends on its focus and settings. Once you submit the prompt, Claude Code processes it under its own configuration and policies.

If native /voice already covers your language and environment, start there. For Chinese-English instructions across Mac apps, try Ducduc with the short phrase above before dictating a full task. For native recording, submission, quota, and SSH limits, use the separate Claude Code voice limits guide.