Caw for Android

Caw is hold-to-talk dictation for Android. Hold Crumb (the little 3D meatball on screen) or the volume-down key, speak, let go, and your words appear.

Speech recognition runs entirely on the phone with sherpa-onnx and the Moonshine base English model. It needs no Google services and no network to dictate.

Caw was made by someone whose speech standard dictation doesn’t follow well: an accent, a mumble, a lot of stopping and starting. It’s for anyone whose speech gets lost the same way, including people who stutter. Pauses, restarts and repeated sounds are expected, not treated as mistakes, and a hold can last up to 5 minutes.

What it does

• Dictation app. Opening Caw shows the dictation screen: a transcript of everything you said, with Copy and Clear (Clear can be undone for a few seconds).
• Voice input method. Caw is also an Android input method with a voice subtype, so a keyboard's mic key can hand off to it. With HeliBoard, tap the mic key, hold Crumb, let go: the text goes into the text box and Caw switches back to HeliBoard.
• Send to computer. Paired with a computer running Caw Desktop (Linux or Windows, a separate download), Caw can type your dictation on the computer instead, with Backspace and Enter buttons. Pairing is by QR code or a typed code; the connection is TLS 1.3 on your local network, with both sides pinned by certificate fingerprint. The phone remembers every computer you pair; you pick the one to send to on the Computer screen.
• Touchpad. The phone can move the paired computer's mouse: one finger moves, tap clicks, two fingers scroll, a two-finger tap right-clicks, press and hold to drag.
• Word fixes. Every recognizer mishears some of your words the same way every time. Caw keeps a list of "when it hears this, write that" in word-fixes.json in the app's files (the same format as Caw for Mac). It ships empty. To build in your own list, put it at ~/.config/caw-android/assets/word-fixes.personal.json on the computer you build on; that folder is outside the repo, so your list never ends up in git.
• Settings. When the phone is paired with a computer, a Settings screen shows the computer's word fixes (add, edit, delete) and its touchpad pointer speed and acceleration, plus a phone-only natural-scrolling switch.
• A built-in cleaner removes fillers and stutters and tidies spacing and capitals, and Caw tells you when the mic seems covered ("I can't hear you").

Private by design

• Dictation never leaves the phone. The only network use is Send to computer and the touchpad, and only to a computer you paired.
• No accounts, no analytics, no Google Play services.
• Crumb's page is served from the app's own files; the WebView blocks every other request.
• Caw never saves your recordings. Audio is turned into text and thrown away.

Install

You don't have to build Caw to use it.

1. On the phone, open the latest release and download caw-android-<version>.apk (about 310 MB, most of it the speech model).
2. Open the downloaded file. Android asks you to allow installs from your browser or file manager the first time; allow it, then tap Install.
3. Open Caw and hold Crumb.

Or from a computer, with the phone connected over USB:

adb install caw-android-0.2.apk

Every release is signed with the same key. Its certificate's SHA-256 fingerprint is

25:4E:55:52:44:E3:B7:E3:FE:AA:12:DA:8E:D5:31:CB:17:8D:AB:FC:A4:12:46:F5:69:96:05:C1:6C:A2:C4:5F

(apksigner verify --print-certs caw-android-0.2.apk shows it). Each release installs over the one before and keeps your settings and pairings. A copy of Caw you built yourself is signed differently: uninstall it first.

Requirements

• An arm64 Android phone running Android 10 (API 29) or later. The APK only carries arm64 libraries.
• To build: JDK 17 and the Android SDK (platform 35). Android Studio is not needed.
• About 300 MB free on the phone: the speech model is packed into the APK.

Build and run

1. Point Gradle at your Android SDK: set ANDROID_HOME, or write sdk.dir=/path/to/android-sdk into local.properties (not tracked).

2. Get the sherpa-onnx Android library. It is a GitHub release asset, not on Maven Central:

   curl -fL -o app/libs/sherpa-onnx.aar \
     https://github.com/k2-fsa/sherpa-onnx/releases/download/v1.13.8/sherpa-onnx-1.13.8.aar
   

3. Get the speech model (about 290 MB, not tracked):

   scripts/fetch-model.sh
   

   It downloads sherpa-onnx-moonshine-base-en-int8 from the sherpa-onnx releases into app/src/main/assets/models/moonshine-base-en/. Without it the app builds but cannot transcribe. (scripts/fetch-model.sh streaming fetches an older streaming model that Caw falls back to when Moonshine is absent.)

4. Build, test and install (phone connected, adb devices shows it):

   ./gradlew testDebugUnitTest
   ./gradlew installDebug
   

5. Open Caw and hold Crumb. Caw asks for the microphone the first time (and, on Android 13+, for notifications: a silent "Caw is ready" notification keeps the speech model loaded in the background).

To use Caw from a keyboard, turn it on in Settings › System › Keyboard › On-screen keyboard › Manage keyboards (the wording varies by phone), then tap HeliBoard's mic key. If the mic key does nothing, make sure only one voice input method is enabled; HeliBoard uses the first one it finds.

scripts/release.sh builds the signed release APK. It needs the release key, which lives outside the repo, so it only works for the maintainer.

scripts/caw_desktop_stub.py is a small test desktop for Send to computer and the touchpad: it prints what the phone sends and answers like Caw Desktop would (Python 3.8+ with TLS 1.3, standard library only).

Status

Personal project. It works for the author; no support is promised.

Licence

• Code: MIT, see LICENSE.
• Artwork, the Crumb character, and the names Caw and Crumb: all rights reserved, see ART-LICENSE.md.
• Third-party code and models (sherpa-onnx, ONNX Runtime, Moonshine, three.js, ZXing, AndroidX, Kotlin): see THIRD_PARTY_LICENSES.md.