The Flutter app is Local Whisper’s mobile speech-to-text surface. It gives iOS and Android a recorder app, local model packs, searchable history, modes, and a native keyboard for text fields. Record in the app, keep data on the device, and use modes to shape the finished text. The iOS keyboard extension and Android input method bring Local Whisper actions into other apps. The iOS app requires iOS 15 or later. Apple SpeechTranscriber requires supported iOS 26 hardware.

​
Status

​
First-run flow

First launch shows setup before the tab shell:
  1. Welcome
  2. Recommended model install
  3. Microphone permission
  4. Keyboard setup and practice
  5. Finish
You can replay setup from Settings. The progress indicator is read-only; navigation happens through explicit actions. Optional model choices open in place.

​
Architecture

Flutter owns the app shell, local model-pack management, local history, modes, settings, clipboard output, and deterministic cleanup. Native iOS uses:
  • ios/Runner/LocalSpeechBridge.swift for microphone recording, WhisperKit/Core ML, and Apple SpeechTranscriber.
  • LocalWhisperUI/Sources/AppleSpeechCore/AppleSpeechCore.swift for the shared macOS/iOS SpeechAnalyzer implementation.
  • ios/LocalWhisperKeyboard/ for mode buttons, punctuation, haptics, and Verify.
Native Android uses:
  • MainActivity.kt for microphone status, 16 kHz mono WAV recording, app settings, input-method settings, keyboard status, and preference sync.
  • LocalWhisperInputMethodService.kt for Verify, mode markers, punctuation, Space, New line, Delete, settings, and haptics.
  • AndroidManifest.xml for microphone, haptics, app identity, launcher identity, and input-method service.
  • lib/src/sherpa_speech_service.dart for Flutter-side sherpa-onnx transcription in a background isolate.

​
History and keyboard control

History stays on the device. The History tab shows saved entry count, total recorded duration, and approximate saved-history size. You can export history as Markdown to the clipboard, delete one transcript, or clear all history after confirmation. The Models tab shows installed pack count and model storage. Installed packs can be removed from the device when you no longer need them. Custom modes use deterministic offline formatting. Instructions containing bullets, notes, or list select bullet formatting; concise, message, chat, or single-line instructions flatten line breaks; task, prompt, or instruction adds a task prefix. Other instructions keep the cleaned text unchanged. Edit or delete custom modes from the Modes tab. The iOS keyboard reads haptics and quick-insert settings from the shared app group. The Android keyboard reads the same settings from setup preferences. Turning off quick insert keeps Verify, Space, New line, and Delete available while hiding mode and punctuation shortcuts.

​
Model packs

Only the wired packs produce transcripts on the device today. The catalog entries install and verify, and the app labels them as needing a native adapter. Large model-pack installs retry transient manifest and file-download failures. Downloaded files are verified before a pack is marked ready.

​
Android notes

Android debug QA can seed the recommended pack and interaction data:
Android requests microphone permission, records local WAV audio, shows levels, stores data locally, verifies the native input method, and transcribes with the installed sherpa-onnx model pack. Debug QA seeds interaction state so emulator passes can exercise the app and keyboard flow without downloading a large model each time. Large model downloads can fail on flaky networks or temporary Hugging Face edge responses. The app retries transient manifest and file failures automatically. If the install still fails, tap Download again after the connection is stable.

​
Checks

Run from src/flutter/local_whisper:
Download and verify the selected pack as part of a native transcription test on an iOS simulator or Android emulator:
The test selects WhisperKit on iOS and Parakeet INT8 ONNX on Android. Model downloads need network access; speech inference uses the installed local pack. --no-uninstall keeps the pack after the test. To reuse an installed pack, pass its path inside the running app’s current container:
To exercise Apple SpeechTranscriber, use a supported physical iOS 26 device. Simulators can report the runtime unavailable, which the test treats as an explicit capability result: