EchoMind is a React Native (Expo Router) Android app for fully on-device voice intelligence.
It combines:
- Whisper for local speech-to-text
- Gemma 3 1B (GGUF via
llama.rn) for local chat + memory extraction - A lightweight memory system (Notes / Tasks / Facts) used as conversational context
- Records live microphone audio and transcribes it locally.
- Lets the user upload an audio file and transcribe it locally.
- Extracts useful memory items from transcripts using Gemma.
- Uses relevant memory items as context for chat answers.
No cloud inference is required for the main AI pipeline.
- Expo SDK 54 + React Native 0.81
- Expo Router (tabs-based app)
whisper.rn(on-device Whisper + VAD)llama.rn(GGUF model runtime)- TypeScript
gen2026/
app/
(tabs)/
index.tsx # Record + realtime transcript + file upload transcription
memory.tsx # Memory management UI (Notes / Tasks / Facts)
chat.tsx # Chat UI powered by Gemma + relevant memory context
services/
whisper-service.ts # Whisper model handling, WAV processing, live/file transcription
llm-service.ts # Gemma download/load/inference + memory extraction
memory-store.ts # In-memory store for notes/tasks/facts
background-service.ts # Android foreground notification for live transcription
native-runtime.ts # Native runtime guards
scripts/
run-android-dev.ps1 # Recommended Windows Android dev flow
gen2026_vendor_backup/
bitnet.cpp/ # Backup/vendor folder
- Node.js 18+
- npm
- Android Studio + Android SDK
- A connected Android device (or emulator)
This app depends on native modules (whisper.rn, llama.rn).
Expo Go is not supported for inference features. Use an Android development build (dev client).
From the repository root:
cd gen2026
npm run android:devclientThis script will:
- mirror the project to
C:\g26(avoids Windows path-length issues) - install dependencies
- run Expo prebuild if needed
- build/install Android debug APK
- start Metro in dev-client mode on port
8083 - launch the app on your connected Android device
cd gen2026
npm install
npx expo start --dev-client --host lan --port 8083For a full native rebuild in the short path workspace:
cd C:\g26
npm install
npx expo prebuild --clean --platform android
cd android
.\gradlew.bat installDebug- Tap the mic button to start/stop live transcription.
- Watch realtime transcript in the transcript panel.
- Optionally attach an audio file and press Transcribe.
- Review/edit memory entries grouped as:
- Notes
- Tasks
- Facts
- Memory can be manually edited and AI-populated from transcripts.
- Chat with the local Gemma model.
- The app injects relevant memory context into prompts.
- Whisper model files are downloaded and cached on device.
- Gemma model file:
gemma-3-1b-it-IQ4_NL.gguf(local device storage).
If Gemma load fails due to stale/corrupt cache, clear app storage or reinstall the app.
- Fixed via keyboard-avoiding layout in Chat and Memory tabs.
- Clear app data / reinstall app to remove cached model and re-download.
- Ensure you are using a native Android dev build (not Expo Go).
- Usually indicates stale binary.
- Reinstall app and rebuild dev client.
Run these inside gen2026/:
npm run start— start Exponpm run android— run Android appnpm run android:devclient— recommended Android dev workflownpm run web— start web build (UI only; native inference unavailable)npm run lint— lint project
EchoMind demonstrates an on-device AI assistant pipeline:
- speech input → local transcription → memory extraction → context-aware local chat
with a mobile UX focused on privacy and low-latency iteration.