Voice-controlled task manager — Create, read, update, and delete tasks entirely through natural spoken conversation. No typing. No buttons. Just talk.
| Feature | Details |
|---|---|
| Voice CRUD | Create, read, update, delete tasks by speaking |
| Context Memory | "Move the previous one to tomorrow" — understands references |
| Temporal Intelligence | Understands "morning", "evening", "next week", "tomorrow" |
| Interruption Handling | Tap orb mid-speech to immediately interrupt |
| Multi-task Creation | "Create 3 tasks: gym at 7, sync at 9, LinkedIn at 11" |
| Auto-reconnect | WebSocket reconnects automatically with exponential backoff |
| Auth (optional) | JWT signup/login, or use guest mode |
| Flexible Storage | SQLite (dev) or PostgreSQL (Supabase/Render) |
Browser Mic (Web Speech API)
↓
STT Transcript
↓ WebSocket VOICE_INPUT
FastAPI WebSocket Handler
↓
ConversationService ←→ ContextService (session memory)
↓
LLMService (Groq llama-3.3-70b) → structured intent JSON
↓
TaskService (CRUD + smart matching)
↓ WebSocket AI_RESPONSE + TASK_UPDATE
Browser TTS (SpeechSynthesis API)
- Python 3.10+
- Node.js 18+
- A free Groq API key (takes 30 seconds)
cd tasktalk-ai/backend
# Copy env
copy .env.example .env
# Edit .env and set your GROQ_API_KEY
# Install dependencies
pip install -r requirements.txt
# Start backend
uvicorn app.main:app --reload --port 8000cd tasktalk-ai/frontend
npm install
npm run devVisit http://localhost:5173 and tap the orb
Note: Voice requires HTTPS in production. For local dev, Chrome allows mic on localhost.
| Variable | Required | Description |
|---|---|---|
GROQ_API_KEY |
✅ | Get free at console.groq.com |
DATABASE_URL |
✅ | SQLite (default) or PostgreSQL URL |
SECRET_KEY |
✅ | JWT secret (32+ chars) |
FRONTEND_URL |
✅ | CORS allowed origin |
| Variable | Default | Description |
|---|---|---|
VITE_API_URL |
http://localhost:8000/api/v1 |
Backend REST URL |
VITE_WS_URL |
ws://localhost:8000/ws |
Backend WebSocket URL |
"Create a task for gym tomorrow at 7 AM"
"Create three tasks: gym at 7, team sync at 9, and LinkedIn post at 11"
"What are my morning tasks?"
"Give me today's agenda"
"Move the LinkedIn task to 6 PM"
"Actually change the previous one to tomorrow"
"Delete the 9:15 task"
"Yes" (confirms deletion)
"No" / "Cancel" (cancels)
| Layer | Technology |
|---|---|
| Frontend | React 18 + Vite + Vanilla CSS |
| State | Zustand |
| STT | Web Speech API (browser-native) |
| TTS | SpeechSynthesis API (browser-native) |
| WebSocket | FastAPI WebSockets |
| LLM | Groq API (llama-3.3-70b-versatile) |
| Backend | FastAPI + SQLAlchemy |
| Database | SQLite (dev) / PostgreSQL (prod) |
| Auth | JWT + bcrypt |
| Frontend Deploy | Vercel |
| Backend Deploy | Render |
tasktalk-ai/
├── backend/
│ ├── app/
│ │ ├── main.py # FastAPI app entry point
│ │ ├── database.py # SQLAlchemy async engine
│ │ ├── config.py # Settings via pydantic-settings
│ │ ├── models/ # ORM models (Task, User)
│ │ ├── schemas/ # Pydantic schemas (intent, auth, task)
│ │ ├── services/
│ │ │ ├── llm_service.py # Groq intent extraction + response gen
│ │ │ ├── task_service.py # CRUD + smart matching
│ │ │ ├── conversation_service.py # Full pipeline orchestration
│ │ │ ├── context_service.py # Session memory (TTL-managed)
│ │ │ └── auth_service.py # JWT + bcrypt auth
│ │ ├── websocket/handler.py # WS endpoint with interrupt support
│ │ ├── routers/auth.py # REST auth endpoints
│ │ ├── prompts/system_prompt.py # LLM system prompt
│ │ └── utils/temporal.py # Natural language date/time parsing
│ ├── requirements.txt
│ └── .env.example
│
├── frontend/
│ ├── src/
│ │ ├── App.jsx # Main layout + auth gate
│ │ ├── main.jsx # React entry
│ │ ├── index.css # Complete design system
│ │ ├── store/useStore.js # Zustand global state
│ │ ├── services/ws.js # WebSocket client with reconnect
│ │ ├── hooks/
│ │ │ ├── useWebSocket.js # WS event → store wiring
│ │ │ ├── useVoice.js # Web Speech API + VAD
│ │ │ └── useTTS.js # SpeechSynthesis + interrupt
│ │ └── components/
│ │ ├── VoiceOrb.jsx # Animated 5-state orb
│ │ ├── TranscriptPanel.jsx # Conversation history
│ │ ├── TaskPanel.jsx # Live task sidebar
│ │ ├── StatusBar.jsx # WS status + latency
│ │ └── Auth.jsx # Login/Signup/Guest
│ └── package.json
│
├── render.yaml # Render backend deploy config
└── README.md
cd frontend
npm run build
# Push to GitHub, connect repo to Vercel
# Set env vars: VITE_API_URL, VITE_WS_URL- Push repo to GitHub
- Create new Web Service on render.com
- Set root directory:
backend - Build command:
pip install -r requirements.txt - Start command:
uvicorn app.main:app --host 0.0.0.0 --port $PORT - Add environment variables from
.env.example
DATABASE_URL=postgresql+psycopg2://user:password@db.supabase.co:5432/postgres
- Create tasks: "Create a task for team sync tomorrow at 10 AM"
- Read agenda: "What are my tasks for tomorrow?"
- Update by reference: "Move the second one to 11 AM"
- Context update: "Actually change it to the day after tomorrow"
- Interrupt: While assistant speaks, tap orb → stops immediately
- Multi-task: "Create three tasks: gym at 7, sync at 9, LinkedIn at 11"
- Delete with confirm: "Delete the gym task" → "Yes"
- Semantic understanding: "Cancel my morning workout" (finds gym task)
- Web Speech API requires Chrome or Edge (not Firefox/Safari)
- Voice recognition quality depends on microphone and browser
- Groq free tier has rate limits (~30 req/min) — sufficient for demos
- SQLite does not support concurrent writes (use PostgreSQL for production)
- Recurring tasks ("every Monday")
- Priority levels via voice
- Calendar export (iCal)
- Deepgram STT integration for higher accuracy
- Push notification reminders
- Multi-language support