Android app that transcribes voice messages on-device with whisper.cpp, wired into the share sheet. Tested mostly on a Pixel 9.
- Local only. Audio never leaves the phone.
- Share to transcribe. Long-press a voice note in WhatsApp / Telegram / Signal, Share, pick WhisperShare.
- GPU path via Vulkan (Mali-G715). CPU is fast enough on small/base that I rarely flip it on.
- Multilingual. German, English, etc., or auto-detect. Optional translate-to-English.
Audio never leaves your device. The only network call is the opt-in model
download from huggingface.co when you tap the cloud icon. No analytics, no
crash reporters, no third-party SDKs. INTERNET is the only permission.
Full policy: https://daniilbabanin.github.io/WhisperShare/privacy/
- GitHub release. Sideload
WhisperShare.apkfrom the latest release. - Obtainium. Add
https://github.com/DaniilBabanin/WhisperShareas a source.
arm64-v8a, Android 12+.
- Launch once. Tap the cloud icon next to Base (multilingual) to download the default model (~60 MB), then select it.
- In WhatsApp/Telegram/Signal, long-press a voice message and share to WhisperShare.
- Text appears with Copy / Share buttons.
Settings:
- Use GPU when available. Only does anything if you built with
whispershare.vulkan=true. - Language. Empty for auto-detect, or
de/en/fretc. - Translate to English. German/Russian/etc. audio comes out as English text.
Android Studio Ladybug (2024.2)+ with NDK side-by-side.
git clone git@github.com:DaniilBabanin/WhisperShare.git
cd WhisperShare
# Android Studio will prompt for: SDK 35, NDK 27.x, CMake 3.22.1.
# For Vulkan: set whispershare.vulkan=true in gradle.properties (+~3 MB APK).CLI:
./gradlew assembleRelease # app/build/outputs/apk/release/app-release-unsigned.apk
./gradlew installDebug # debug build straight to deviceFrom huggingface.co/ggerganov/whisper.cpp:
| Model | Size | Speed on Pixel 9 (CPU) | Quality |
|---|---|---|---|
tiny-q5_1 |
31 MB | ~15× realtime | OK for clear English |
base-q5_1 |
60 MB | ~8× realtime | Recommended default |
small-q5_1 |
190 MB | ~3× realtime | Best for accents / noise / mixed languages |
base.en-q5_1 |
60 MB | ~10× realtime | English-only, slightly faster than base-q5_1 |
For German, use base-q5_1 or small-q5_1. With Vulkan, small runs ~1.5–2×
faster than CPU; on tiny/base, dispatch overhead eats the win.
[Share intent] ─▶ TranscribeActivity
│
▼
AudioDecoder (MediaCodec)
decodes OPUS/AAC/MP3 → PCM 16 kHz mono float
│
▼
WhisperEngine (JNI → whisper.cpp)
runs whisper_full() in a coroutine
│
▼
TranscribeScreen
(selectable text + Copy/Share)
libwhispershare.so is built by CMake during gradle sync. FetchContent pulls
whisper.cpp v1.7.4 the first time.
app/
├── build.gradle.kts # NDK ABIs (arm64-v8a only), Compose, CMake glue
├── CMakeLists.txt # Fetches whisper.cpp, builds JNI .so
├── src/main/
│ ├── AndroidManifest.xml # Share intent filters
│ ├── cpp/whisper_jni.cpp # JNI bridge to whisper_full()
│ ├── kotlin/io/whispershare/
│ │ ├── WhisperApp.kt
│ │ ├── WhisperEngine.kt # Kotlin wrapper of JNI calls
│ │ ├── AudioDecoder.kt # MediaCodec → 16 kHz float PCM
│ │ ├── ModelManager.kt # Download & list ggml models
│ │ ├── AppPreferences.kt
│ │ ├── MainActivity.kt # Home / settings
│ │ ├── TranscribeActivity.kt # Share-target activity
│ │ ├── ShareUtils.kt
│ │ └── ui/ # Compose screens + theme
│ └── res/ # Strings, themes, launcher icon
└── proguard-rules.pro
- Cache transcriptions in Room (currently they vanish on close).
- Foreground service for files >5 min.
- VAD pre-trimming with
ggml-silero-vad. - Notification with the text so the app doesn't need to stay foregrounded.
- Translate to languages other than English.
- First fresh build is slow. whisper.cpp gets cloned and compiled for arm64. Incremental builds are quick.
- Vulkan on Mali can crash on very old quantizations. Set
whispershare.vulkan=falseand rebuild if it does.
MIT