Independent offline Omi companion for Android 8+. Public Android application source, tests and build tooling. Not affiliated with Omi or Based Hardware.
Explore the app and full feature guide →
Android only for now — no iOS version. Latest signed download: 0.4.13 / build 20, an Android preview for Android 8+ (API 26+) arm64-v8a phones. This is a prerelease, not a stable release.
Download v0.4.13 APK ↓ · Release notes, checksums and licences
The feature guide covers the downloadable 0.4.13 APK, including the profiles, smart actions, flashlight, four/five-press controls and notification app/pattern choices introduced in 0.4.12 source. Version 0.4.13 adds a notification filter choice and allows screenshots throughout the app.
Source version: 0.4.13 / versionCode 20. Application license: MIT; third-party components retain their own notices.
- This branch contains the 0.4.13 source, including the earlier capture/storage features and all controls described below.
- Latest v0.4.13 preview APK is available for
arm64-v8a, with matching checksums, manifest and licences attached to that release. It retains the 15 GB PCM quota, RF64 export and queue-capacity recovery alongside the expanded controls, notification filters and screenshot support. - Older v0.4.11 preview remains available as a historical release. The 0.4.12 checkpoint was published as source, without a public APK release.
- Older v0.4.7 preview remains available as a historical release.
- The v0.3.1 download, root
SHA256SUMS.txt,release-manifest.jsonandRELEASE-NOTES-0.3.1.mddescribe older 0.3.1 binaries, not the 0.4.13 release. - A source ZIP from the current branch contains the 0.4.13 application checkpoint, not an installable APK. Automatic source archives attached to older release tags remain historical snapshots; publishing this branch does not change those tags or release assets.
- Package:
app.nottheomi.ai. Install updates over the existing app only with the same signer; do not uninstall or clear data to preserve the local encryption key and recordings. Your own signing key cannot update an existing differently signed installation.
The implementation notes and build instructions below describe 0.4.13 source. The feature guide covers the same feature set as the v0.4.13 APK.
- Omi Bluetooth LE Opus audio → encrypted phone-local PCM and transcript history.
- Bundled Vosk for streaming English drafts; CPU-only Whisper small.en Q5_1 for post-save refinement.
- Searchable calendar/day Library, bounded previews, playback and explicit WAV/RF64/text export, with a 15 GB retained PCM quota.
- Independent single/double/triple/four/five-press actions and manually selected profiles, plus button-only standby without recording audio.
- Optional notification vibrations with independent notification filtering, app selection and default/per-app patterns.
- No Omi account, PC relay, cloud transcription, runtime model download or
INTERNETpermission. - Capability-gated battery, brightness and button controls. Long-press/power behavior stays firmware-owned.
- Explicit phone-microphone fallback; no automatic substitution when the wearable is absent.
- Disconnect other Omi apps, wake/charge the wearable and enable Bluetooth.
- Tap Connect Omi, grant Nearby devices and notifications, then select your device. Android 8–11 also needs Location permission and system Location for BLE discovery; the app does not collect location.
- Recording starts when decoded Omi audio arrives. Preparing/Connecting is not Recording. Initial model verification/copying is local.
- Stop & save ends recording and retains a button-only connection. Saved sessions remain in Library; Whisper refinement runs after capture stops. Disconnect buttons closes the remaining connection.
App launch/reboot never auto-records. Cancellation before the first audio and fatal failures terminate rather than reopening standby. Brightness zero requests minimum, not guaranteed darkness. Optional controls stay unavailable when characteristic/readback validation fails.
In Omi device & controls, manually choose Everyday, Meeting or Music. Each profile retains independent actions and shortcut targets. Everyday preserves existing mappings; Music starts with play/pause and next track. Switching profiles cancels pending gestures; notification preferences remain shared. Profiles do not switch automatically.
Each profile has independent Single press, Double press, Triple press, Four presses and Five presses dropdowns. Available actions:
- Bookmark in transcript (original single-press default).
- Start recording.
- Stop & save (original double-press default).
- Start / stop recording.
- Play / pause music.
- Pause music.
- Play music, Next track, Previous track, Stop music, Fast-forward music and Rewind music.
- Media volume up, Media volume down and Mute / unmute media (media stream only).
- Phone shortcut.
- Flashlight toggle.
- Smart: bookmark recording / play-pause idle.
- Pause music → start recording.
- No action.
Four/five start with No action. Triple-only retains its 650 ms fixed grouping window while four/five remain unassigned. Assigning four extends the window to 950 ms; assigning five extends it to 1250 ms. The highest assigned count sets the window, measured from the first received firmware tap event rather than a sliding deadline. Shorter gestures wait for that window, so enabling higher counts adds latency. Nearby separate gestures can combine. With triple/four/five all unassigned, original single/double timing is retained. This is experimental app-side grouping of stock single/double notifications, not native firmware gestures.
Smart: bookmark recording / play-pause idle inserts a bookmark during recording and requests music play/pause in controls-only standby. Pause music → start recording is a bounded built-in action: it requests Pause, then starts only when idle; it never stops an existing recording. Android does not acknowledge whether the media player actually paused. No arbitrary macros or automatic profiles are provided.
Flashlight toggle uses Android CameraManager, not a camera preview or photo capture. Use Allow flashlight permission to explicitly grant the optional Camera permission; selecting an action does not grant permission or turn on the torch. Status follows observed torch callbacks and reports unknown, unsupported or camera-busy states rather than claiming success. Background and locked-screen behaviour depends on Android and the phone.
Use Connect for buttons — no recording to start button-only standby directly. The app does not subscribe to the audio stream or create a recording in this mode. A configured start/toggle press can transition into recording; Stop & save drains accepted audio, finishes the recording, then returns to standby. Disconnect ends both capture and button control.
Music functions dispatch Android media keys to the active media session. Shortcut selection supports eligible exported phone activities and supported shortcut providers; background/locked-screen launches use a tap-to-run notification when required by Android. These are app-local mappings while connected, not firmware remapping or unrestricted phone automation.
Real four/five-press cadence, motor patterns and locked-screen torch physical acceptance remain pending. Media-app compatibility, shortcut delivery and sustained wearable reliability also need hardware acceptance; synthetic tests do not certify these behaviours.
Notification vibrations remain off by default and opt-in, requiring Android notification access and an existing recording or button-only Omi connection. Upgrading preserves the previous opt-in state, selected packages and pattern overrides; All apps and Double pulse remain the defaults for those separate settings.
- Notification filter is separate from app selection. Choose All notifications (the default when no filter preference is saved) or Alerting only.
- All notifications includes silent/low-importance, grouped (including group-summary), ongoing, foreground-service and DND-delivered posts.
ONLY_ALERT_ONCEis not a blanket exclusion. This mode can request Omi motor pulses even during DND (Do Not Disturb); Android must still deliver the post to the listener. - Alerting only retains restrictive filtering: silent/low-importance, ongoing/foreground-service and group-summary posts are skipped, and automatic alerts are suppressed during DND.
- Choose All apps or Selected apps only. An empty selected-app list produces no notification pulses.
- Choose a default Short pulse, Double pulse, Triple pulse or Long pulse, with per-app overrides. Overrides apply in either app-selection mode.
- Notification settings are independent of the manually selected control profile. Changing notification choices cancels queued vibrations; it does not grant access, connect or record.
- Both filters retain own-app and package exclusions, stale and duplicate suppression, the 1.5 seconds burst limit, busy-write safeguards and reconnect fences. An eligible post does not guarantee a pulse. Message text is not read, saved or sent to Omi; filtering uses metadata only.
- The manual test does not itself check DND. It uses the selected default pattern on the existing connection. BLE acknowledgments are not measurements of motor output.
- Patterns use the existing BLE owner: no second BLE connection, no replay backlog or phone-motor substitution. WhatsApp physical motor delivery is not verified; motor-pattern acceptance remains pending.
Initial startup without decoded audio is bounded at 120 seconds. After an explicit session has received PCM, prolonged audio absence remains Waiting for Omi with capped-backoff recovery until explicit Stop/Disconnect or a terminal failure. Connection callbacks, battery traffic and malformed packets do not prove recovery.
Reported BLE gaps create separate encrypted continuation recordings when actual audio resumes. Playback/export do not splice audio across a known gap; retries without decoded PCM do not create empty entries. Accepted pre-Stop audio still drains to storage. Missing audio is not reconstructed, and pauses without a transport gap may remain undetected.
Storage and live-preview queues are independently bounded. Audio is archived before preview recognition. Whisper refines saved encrypted PCM in bounded batches with durable encrypted checkpoints; interrupted work resumes from its last committed offset. A permanently blocked native recognizer can delay final save/teardown. Real screen-off and workday endurance remain unverified.
AES-GCM audio/text with an Android Keystore key; app backups are disabled. Screenshots are allowed throughout the app, including screens with transcripts or settings. Screenshots can include visible private content; review them before sharing. Uninstall/data clear loses the key. User-requested exports are plaintext. Device selection and controls stay on the phone; transcript text is not placed in notifications/logs.
The 15 GB aggregate retained PCM quota (15,000,000,000 bytes) includes existing recordings; 128 MiB free-space reserve. Large WAV exports use RF64; models, transcripts, database overhead and exports use additional space. English only. Recognition may be wrong. Process death may lose an uncommitted tail. Record with participants' permission.
Prerequisites: Java 17, Python 3, Android SDK/platform 34, Android build-tools 34.0.0, Android NDK 27.2.12479018, and CMake. Native build scripts target Linux/WSL. Set ANDROID_SDK_ROOT to your SDK location and configure sdk.dir in your local, untracked local.properties if needed.
python3 scripts/prepare_model.py
python3 scripts/build_whisper.py --ndk "$ANDROID_SDK_ROOT/ndk/27.2.12479018"
./gradlew --no-daemon assembleDebug assembleRelease assembleDebugAndroidTest lintDebug --console=plainPreparation downloads and verifies pinned public models/source. Native compilation itself does not download. Provenance is pinned in DEPENDENCIES.json. Supported ABIs: arm64-v8a, x86_64; no 32-bit ARM. Model weights, generated native libraries, APKs, signing keys and machine configuration are excluded from Git.
Use python3 scripts/sign_release.py --help for signing with your own external key directory. Original release signing keys are not published. tests/hybrid/verify_apk.py checks models/native payloads, both speech engines, signature and lack of Internet permission.
See the 0.4.13 public-source verification scope for publication checks, retained candidate evidence and remaining physical gates.
Retained 0.4.13 evidence includes 112 targeted signed-emulator tests and a separate 11-test ad-hoc rerun. The rerun is non-additive; this is not a full-suite result or hardware acceptance, and Android tests were not freshly rerun for publication. A 0.4.12 → 0.4.13 USB update was verified against the exact APK bytes, installed identity and permissions. No recording was started during that installation check. Real four/five cadence, motor patterns, WhatsApp delivery and locked-screen torch checks remain pending.
Historical 0.4.12 evidence includes 125 Android instrumentation tests, 54 gesture scenarios, and 641 haptic/BLE scenarios. These checks overlap; do not add the counts together or treat them as current 0.4.13 results. The 0.4.12 verification scope, 0.4.7 verification scope and earlier documents remain historical records.
The documentation regression is independent of Android builds and model preparation:
python3 tests/pages/check_site.pyNo phone is needed for these controlled host checks. First prepare pinned model assets and configure sdk.dir: capture installer checks use the real models and Android SDK; other host entrypoints use controlled doubles.
python3 tests/omi/run_ble_tests.py
python3 tests/omi-capture/run_host_checks.py
python3 tests/omi-composition/run_host_checks.py
python3 tests/capture/run_host_checks.py
python3 tests/hybrid/run_host_checks.py
python3 tests/hybrid/run_job_checks.py
python3 tests/whisper-java/run_host_checks.py
python3 tests/library-dates/run_host_checks.py
python3 tests/buttons/test_android_runner.pyAndroid instrumentation needs prepared artifacts and an explicitly selected synthetic target. tests/buttons/run_android_checks.py is constrained to its isolated emulator and rejects an active app service; it is not a production-phone installer. KVM-backed x86_64 execution is recommended: an older software-emulation path produced an instruction overread absent under KVM with unchanged APK bytes.
For tests/whisper-device/run_probe.py with Windows adb.exe, pass --windows-temp with an existing Windows-accessible WSL directory. Public jfk.wav and test.wav are upstream fixtures, not private recordings. Production-phone diagnostics, private transcripts, signing material and private Git history are excluded.