VocaPhone
Voice dictation keyboard that transcribes on hardware you control
新版本 0.2.5
Cancel dictation while you are still speaking: the keyboard X now shows during Listening. Leaving a field mid-dictation saves the transcript to History instead of losing it, and failures say what went wrong and offer Retry. Pick a maximum recording length from 1 to 10 minutes (5 by default). Model downloads resume after a dropped connection. Also new: long-press space to switch keyboards, a setting to turn off auto-capitalization, and Traditional Chinese for Taiwan and Hong Kong.
Speak into your phone and the text shows up where you are already typing. VocaPhone is a system keyboard: switch to it in any app, dictate, and the transcript is inserted at your cursor through Android's normal input method interface.
Transcription never touches a cloud speech service. It runs either on your own device or on a machine you already own and administer via a self-hosted gateway. There are no accounts, no analytics SDK, and no subscription. The Play build offers optional anonymous usage reporting, off unless you turn it on, reporting to a server the developers self-host; it never includes your speech, your text, or your gateway address. The F-Droid build has it compiled out entirely.
Two ways to transcribe
- On device. Download a Whisper model once and dictate with no network at all. The speech engine is whisper.cpp, compiled from source as part of the app. Models range from a few tens of megabytes to several gigabytes, and the app recommends one that fits your phone's memory. Where available, the catalog also includes sherpa-onnx models such as Parakeet.
- Through your own gateway. Point the app at VocaGateway running on your Mac, Linux desktop, or home server, and that machine does the work. This is the faster option for large models and lets a phone use hardware it does not have.
You choose the network path to your gateway: a trusted LAN, a private Tailscale network, or HTTPS behind your own reverse proxy. Access uses bearer tokens issued per device, which you can name and revoke individually, and pairing can be done by scanning a QR code.
Dictation
- 54 selectable languages plus automatic detection: English, Mandarin Chinese, Cantonese, Spanish, French, German, Russian, Portuguese, Italian, Dutch, Polish, Ukrainian, Czech, Slovak, Slovenian, Croatian, Serbian, Bulgarian, Romanian, Hungarian, Greek, Danish, Swedish, Norwegian, Finnish, Estonian, Latvian, Lithuanian, Maltese, Catalan, Turkish, Arabic, Hebrew, Persian, Swahili, Japanese, Korean, Vietnamese, Thai, Indonesian, Malay and Filipino, along with Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Urdu, Kannada, Malayalam, Punjabi, Assamese and Nepali
- Four writing styles: Formal, Casual, Very Casual and Excited, with punctuation that follows the script rather than assuming a Latin full stop
- Start, finish, cancel, retry and undo, all from the keyboard
- Automatic microphone routing, or pick an input explicitly, with the microphone in use shown in the app
- Bluetooth headset microphones are supported
Permissions
The keyboard writes through InputConnection and never reads the contents of the field you are typing into. Camera access is requested only when you tap "Scan QR" to pair with a gateway, and the microphone is used only while a dictation is running, with a foreground service notification for the duration.
About this build
The build distributed here is compiled entirely from source. It leaves out the optional sherpa-onnx speech engine, which upstream ships as a prebuilt binary library, so on-device transcription uses the Whisper model catalog. Gateway transcription is unaffected and can use any engine your gateway supports. On supported builds, sherpa-onnx models such as Parakeet are available alongside Whisper.
VocaPhone is part of the Voca family alongside VocaLinux and VocaMac. Report issues on the GitHub issue tracker.
Transcription never touches a cloud speech service. It runs either on your own device or on a machine you already own and administer via a self-hosted gateway. There are no accounts, no analytics SDK, and no subscription. The Play build offers optional anonymous usage reporting, off unless you turn it on, reporting to a server the developers self-host; it never includes your speech, your text, or your gateway address. The F-Droid build has it compiled out entirely.
Two ways to transcribe
- On device. Download a Whisper model once and dictate with no network at all. The speech engine is whisper.cpp, compiled from source as part of the app. Models range from a few tens of megabytes to several gigabytes, and the app recommends one that fits your phone's memory. Where available, the catalog also includes sherpa-onnx models such as Parakeet.
- Through your own gateway. Point the app at VocaGateway running on your Mac, Linux desktop, or home server, and that machine does the work. This is the faster option for large models and lets a phone use hardware it does not have.
You choose the network path to your gateway: a trusted LAN, a private Tailscale network, or HTTPS behind your own reverse proxy. Access uses bearer tokens issued per device, which you can name and revoke individually, and pairing can be done by scanning a QR code.
Dictation
- 54 selectable languages plus automatic detection: English, Mandarin Chinese, Cantonese, Spanish, French, German, Russian, Portuguese, Italian, Dutch, Polish, Ukrainian, Czech, Slovak, Slovenian, Croatian, Serbian, Bulgarian, Romanian, Hungarian, Greek, Danish, Swedish, Norwegian, Finnish, Estonian, Latvian, Lithuanian, Maltese, Catalan, Turkish, Arabic, Hebrew, Persian, Swahili, Japanese, Korean, Vietnamese, Thai, Indonesian, Malay and Filipino, along with Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Urdu, Kannada, Malayalam, Punjabi, Assamese and Nepali
- Four writing styles: Formal, Casual, Very Casual and Excited, with punctuation that follows the script rather than assuming a Latin full stop
- Start, finish, cancel, retry and undo, all from the keyboard
- Automatic microphone routing, or pick an input explicitly, with the microphone in use shown in the app
- Bluetooth headset microphones are supported
Permissions
The keyboard writes through InputConnection and never reads the contents of the field you are typing into. Camera access is requested only when you tap "Scan QR" to pair with a gateway, and the microphone is used only while a dictation is running, with a foreground service notification for the duration.
About this build
The build distributed here is compiled entirely from source. It leaves out the optional sherpa-onnx speech engine, which upstream ships as a prebuilt binary library, so on-device transcription uses the Whisper model catalog. Gateway transcription is unaffected and can use any engine your gateway supports. On supported builds, sherpa-onnx models such as Parakeet are available alongside Whisper.
VocaPhone is part of the Voca family alongside VocaLinux and VocaMac. Report issues on the GitHub issue tracker.
捐赠
版本
尽管下面提供了 APK 安装包的下载选项，但你应该注意，以这种方式安装将不会收到更新通知，这是一种不太安全的下载方式。 我们建议你安装使用 F-Droid 客户端。
下载 F-Droid-
arm64-v8aarmeabi-v7ax86_64该版本需要 Android 13 及以上版本。
此包由原始开发者构建并签名，并保证对应于此源代码 tarball。
-
arm64-v8aarmeabi-v7ax86_64该版本需要 Android 13 及以上版本。
此包由原始开发者构建并签名，并保证对应于此源代码 tarball。
-
arm64-v8aarmeabi-v7ax86_64该版本需要 Android 13 及以上版本。
此包由原始开发者构建并签名，并保证对应于此源代码 tarball。
新版本 0.2.4
Dictation no longer swallows a half-typed word, and names like "Sarah" or "Paris" keep their capitals in every style. On-device models finish faster: Sherpa decodes at pauses while you speak, Fast and Accurate stream again, and Whisper skips slow efficiency cores. Cancelling an on-device transcription is instant and discards the audio. A mic tap that cannot start the microphone now says so, and VocaPhoneX's bubble costs less while you type.
新版本 0.2.2
Android 13 keyboard setup now recognizes VocaPhone. Dictated corrections keep spaces around neighboring words. Automatic microphone routing uses the phone mic with Bluetooth connected; select Bluetooth explicitly for its mic. Dictation notifications clear when capture ends or startup needs repair. Long Whisper recordings preserve speech across 30-second windows. Model loading, setup, and custom words are improved; the catalog adds Qwen3-ASR and Omnilingual.