Dicta

A modern, minimalistic Android dictation app that uses local/edge ASR (Automatic Speech Recognition) models for fully offline transcription. English-only.

Features

• Fully Offline - All speech recognition happens on-device. No internet required after model download.
• Real-time Streaming - See your words transcribed as you speak with low-latency partial results.
• Multiple Model Options - Choose from 4 Moonshine models based on your accuracy/storage needs.
• Recording History - Save and review past transcriptions with audio playback.
• Export Data - Export all recordings as a ZIP file with JSON metadata and audio files.
• Modern UI - Built with Jetpack Compose, Material3, and Material You dynamic color.
• Privacy First - Your voice data never leaves your device.

Screenshots

A visual walkthrough of the app.

Onboarding — pick a Moonshine model

Home — tap mic to start dictating

Listening — live transcription, red stop

Settings — manage installed models

Models

Dicta uses Moonshine Voice by Useful Sensors for speech recognition.

Model	Size (on disk)	Download	WER	Best For
Tiny Streaming	49 MB	32 MB	12%	Quick notes, low storage
Small Streaming	158 MB	100 MB	7.84%	Daily use (Recommended)
Medium Streaming	289 MB	192 MB	6.65%	Maximum real-time accuracy
Base (Offline)	134 MB	102 MB	~10%	File transcription

WER = Word Error Rate (lower is better). Models are downloaded on first launch. You can switch between models in Settings.

Tech Stack

• Language: Kotlin
• UI: Jetpack Compose + Material3 + Material You
• Architecture: MVVM + Clean Architecture
• DI: Hilt
• Database: Room
• Preferences: DataStore
• ASR Engine: Moonshine Voice (on-device, ONNX Runtime)
• Audio: Android AudioRecord API (16kHz mono)

Project Structure

app/src/main/java/com/example/dicta/
├── di/                 # Dependency injection modules
├── data/
│   ├── local/          # Room database
│   ├── repository/     # Repository implementations
│   └── preferences/    # DataStore preferences
├── domain/
│   ├── model/          # Domain models
│   └── repository/     # Repository interfaces
├── asr/
│   └── moonshine/      # Moonshine ASR engine implementation
├── audio/              # Audio recording
├── presentation/
│   ├── home/           # Main recording screen
│   ├── history/        # Recording history
│   ├── settings/       # Model management & export
│   ├── onboarding/     # First-launch model selection
│   ├── navigation/     # Nav host and screen routes
│   └── theme/          # Material3 theme
└── util/               # Utilities

Building

Prerequisites

• Android Studio Ladybug or newer
• JDK 17
• Android SDK 35+
• Physical ARM64 device (no emulator support -- Moonshine requires ARM64)

Build Debug APK

./gradlew assembleDebug

APK will be at: app/build/outputs/apk/debug/app-debug.apk

Build Release APK

./gradlew assembleRelease

Installation

1. Download the latest APK from Releases
2. Enable "Install from unknown sources" if prompted
3. Install and open the app
4. Select a model to download (Small Streaming recommended)
5. Grant microphone permission
6. Start dictating!

Permissions

• RECORD_AUDIO - Required for speech recognition
• INTERNET - Required for initial model download only
• POST_NOTIFICATIONS - For download progress notifications

Export Format

When you export recordings, you get a ZIP file containing:

dicta_export_[timestamp].zip
├── recordings.json      # Metadata for all recordings
├── audio_1_[name].wav   # Audio file for recording 1
├── audio_2_[name].wav   # Audio file for recording 2
└── ...

The recordings.json contains:

{
  "exportedAt": "2025-01-15T10:30:00Z",
  "appVersion": "1.0",
  "recordingCount": 5,
  "recordings": [
    {
      "id": 1,
      "title": "Recording - Jan 15, 10:30 AM",
      "transcription": "Your transcribed text here...",
      "durationMs": 15000,
      "createdAt": "2025-01-15T10:30:00Z",
      "modelUsed": "MOONSHINE_SMALL_STREAMING",
      "audioFile": "audio_1_recording.wav"
    }
  ]
}

Migration from v1.0

Dicta v2.0 replaced the Vosk ASR engine with Moonshine Voice. See docs/vosk-to-moonshine-migration.md for the full migration process, architectural decisions, and what changed.

License

This project is open source. The Moonshine Voice library and models are licensed under the MIT License.

Credits

• Moonshine Voice - On-device speech recognition engine
• Useful Sensors - Moonshine model providers

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.