Capi Voice Input (卡皮语音输入法) has no account system and does not know who you are. By default, what you say, the text it produces and the words you add stay on your device. A few features send content to third-party services. They are all off by default, and nothing is sent until you allow it on an in-app consent screen.

Speech and text

• The speech recognition model is bundled with the app and runs on your device by default. It works in Airplane Mode.
• Recordings are released from memory once transcribed. The app does not save audio files.
• Transcribed text stays on your device unless you use one of the features below.

Your own words

Your word list, saved phrases and quick replies are stored only on your device. They are not uploaded or synced to any server.

Features that send your content to third parties

All of these are off by default. Before the first time any of them sends anything, the app shows a consent screen (in Chinese and English) stating what is sent, who receives it, what it is used for and what you lose if you decline. Nothing is sent unless you tap “Allow”. Audio and text are agreed separately, and each recipient is agreed separately; changing to a different provider asks again. You can withdraw at any time in Settings (设置) → AI Settings (AI 功能设置) → Where your data goes (数据去向); nothing is sent after that. If you decline, the app remains fully usable with on-device speech recognition. On iPhone, the consent screen appears where the feature is used — in the app, or directly inside the keyboard when a feature is used from the keyboard. The Mac and Android versions ask the same questions in their own dialogs.

How the content is collected: audio is the recording you make after pressing the record button; text is what you ask a feature to process — what you just dictated, the text in the current text field (when you use the keyboard's polish, translate or instruction features, the keyboard reads the text around the cursor in that one field), or a message you copied to reply to (Reply suggestions read the clipboard). To improve accuracy, requests may also include a few related words from your word list, a few sentences before the cursor, and your “Let AI know about me” (让 AI 对我有些了解) note, if you wrote one. We do not collect anything in the background, and the keyboard does not read other apps or log keystrokes.

• Enhanced recognition — sends the audio you just recorded to a speech recognition service to get text back: Volcano Engine / ByteDance (openspeech.bytedance.com) or OpenAI (api.openai.com), whichever credentials you entered.
• Face-to-face and Simultaneous translation — send your speech to the service above (or, in real-time mode, stream it to api.openai.com) and the recognized text to your AI provider for translation.
• Polish, translate, reply suggestions, instructions, reading aloud — send the text described above to the AI provider whose key you entered (for example OpenAI, DeepSeek, Kimi or Volcano Engine), directly from your device. If you use a token we issued instead of your own key, the text goes through our relay voice-input-three.vercel.app/api/refine, which only forwards it and does not store it.
• English feedback — when turned on, each sentence you dictate is sent to your AI provider to get an English version or a more natural phrasing. Saved sentences are also sent to break them into parts and to prepare read-aloud audio.
• Local sync — settings and word lists move directly between your own phone and computer on the same Wi-Fi, encrypted with a six-digit pairing code (newer versions also bring Volcano Engine speech credentials from the computer back to the phone). No server is involved.

Requests that carry none of your content

• Public word list updates — a read-only download of a public correction list.
• Update check — asks Apple's App Store whether a new version exists (app identifier and region only).

Keyboard extension

The keyboard needs “Allow Full Access” to read the storage it shares with the app (your words and phrases) and, only if you use the features above, to send the content described there. It does not log, upload or analyze your keystrokes, and it does not go online unless you use one of those features. Without Full Access the keyboard still types.

Third parties

No ads, no analytics, no crash reporting, no third-party SDKs. We never sell or rent your information. Content you choose to send goes only to the providers named above, only to produce the result you asked for. It includes no device identifiers, location or contacts; if you wrote names in your “Let AI know about me” note, that note is included, so write only what you are comfortable sharing with the provider.

Protection by these providers: we only integrate providers that publicly commit not to use API content to train their models, and we require them to protect this content to the same or a greater degree than this policy. Their own terms apply: OpenAI, Volcano Engine, DeepSeek, Kimi. When you use your own key, the content goes directly between you and that provider and never passes through us.

Children

This app is not directed at children and does not knowingly collect information from anyone, including children.

Changes and contact

If features change in a way that affects this policy, we will update this page. If we ever need to collect data, we will tell you in the app and ask for your consent first. Questions: yunfeilu@thebcec.org.uk