Cassini records audio and makes the transcript on your phone. It labels the speakers and can remember the voices of the people that you name. It keeps the audio and the timed words together in one .opus file. Any audio player plays this file, and other Cassini software reads the transcript in it. On compatible arm64 phones with Android 10 or later, Whistle is included and is the initial default: basic transcription needs no model download or network connection. Once Parakeet is installed and verified, new transcriptions default to it; you can still choose Whistle. Parakeet is a recommended download from dist.gocassini.com, and remains the default on other devices. The Parakeet model is Parakeet Ultra by Moondream, a further-trained NVIDIA Parakeet TDT 0.6B v3 for 25 European languages.
Download the latest release · Changelog
Cassini does all speech recognition and speaker analysis on the phone. It does not upload audio, transcripts or voiceprints. It has no accounts, no analytics and no advertising. The app uses the network only to download the models, and only after you agree. Voiceprints stay in the private storage of the app. On Android 9 or later, direct device-to-device transfer copies notes, the catalogue and saved voices when “Move to a new phone” is on (on by default). Cloud backup is not used; downloaded models and unfinished work are not copied. Cassini never writes a voiceprint into a document or an export. PRIVACY.md gives the full privacy policy.
- Android 8.0 (API 26) or later, on a 64-bit ARM phone (
arm64-v8a). The APK also hasx86_64for emulators. - Included Whistle requires Android 10 or later and an arm64 CPU with dot-product support. Settings shows when it is unavailable.
- The optional Parakeet, speaker and voice models need about 1.3 GB free while installing and about 0.8 GB afterwards. Their total download is about 530 MB.
- 4 GB of RAM is a good start. This is not a tested minimum.
These screenshots come from the Android emulator. The audio is a clip of the FLEURS data set (CC BY 4.0), and the names are test names.
| Notes | Read and listen | Transcribing | Name a speaker | Settings | |
|---|---|---|---|---|---|
| Light | ![]() |
![]() |
![]() |
![]() |
![]() |
| Dark | ![]() |
![]() |
![]() |
![]() |
![]() |
git clone --recurse-submodules https://github.com/codemyriad/cassini-android
cd cassini-android
./scripts/setup.sh
./gradlew :app:assembleDebugThe build needs JDK 17 and the Android SDK. CONTRIBUTING.md gives the tools, the tests and the rules for a change.
:core Pure Kotlin/JVM. The verbatim engine plus pure app rules (statuses, search,
trash arithmetic, exports, speaker stats, model plan).
:data Android library. Room catalogue, DataStore settings, file repositories.
:engine Android library. Native wrappers (sherpa-onnx, libcassini_opus), MediaCodec
decoder, microphone capture, TranscriptionJob, ModelInstaller, ProgressTracker.
:ui Android library. Compose design system and stateless screens.
:app Android application. AppGraph, ViewModels, navigation, services, notifications.
:bench Experiments and calibration on a device (com.android.test).
Never in CI.
app -> ui
app -> data -> core
app -> engine -> core
bench -(instruments)-> app
CONTRIBUTING.md: build, tests, style and pull requests.docs/development.md: worktrees, test tiers, device scripts and the pre-release gate.docs/architecture.md: modules, storage, background work and the build.docs/data-formats.md: storage layout, the portable document, exports and the differences from gocassini.docs/native-runtime.md: the sherpa-onnx AAR and libopus.docs/release.md: tags, signing and the F-Droid draft.docs/design/: the features and the visual design.docs/notes/: historical notes of the 0.1.0 prototype.SECURITY.md,PRIVACY.md,CODE_OF_CONDUCT.mdandCHANGELOG.md.
The cassini-format repository gives the specification of the portable format. The desktop pipeline and the web viewer are in gocassini.
Cassini for Android is free software under the GPLv3. THIRD_PARTY_NOTICES.md gives the licenses and sources of the dependencies and the models. We developed Cassini with AI coding assistance.









