Taskify

A fully offline, privacy-first, voice-driven personal task management application.

Taskify captures spoken input via your microphone, transcribes it locally, and automatically extracts and classifies actionable tasks into the Eisenhower Matrix — all without sending a single byte of data outside your device.

---

Table of Contents

1. About
2. Features
3. Privacy
4. Requirements
5. Installation
6. Model Downloads
7. Usage
8. Configuration
9. Architecture Overview
10. Contributing
11. Roadmap
12. License

---

About

Modern task management tools demand an internet connection, a subscription, and trust in a third-party cloud. Taskify is built on the opposite premise: your voice, your hardware, your data — nothing else.

Speak naturally while working, thinking, or planning. Taskify listens, transcribes in near real-time, and periodically runs a local AI pipeline to extract discrete tasks from the transcript. Each task is automatically placed in one of the four Eisenhower Matrix quadrants:

Quadrant	Description
Urgent and Important	Do immediately
Important but Not Urgent	Schedule it
Urgent but Not Important	Delegate or batch
Neither Urgent Nor Important	Eliminate or defer

The application is designed for mid-range consumer hardware (Intel Core i5 / 8 GB RAM or equivalent). It anticipates future Android deployment and is architected to swap in lightweight components with minimal code changes.

---

Features

• 100% offline operation after initial model downloads.
• Voice capture via system microphone with energy-based silence detection.
• Multilingual speech-to-text using Vosk (primary) and faster-whisper (fallback).
• Local AI task extraction via Ollama (Phi-4-mini, Gemma 3 2B, Qwen2.5 quantized) or spaCy/rule-based NLP when a GPU/LLM is unavailable.
• Eisenhower Matrix dashboard with drag-and-drop quadrant management.
• Automatic daily logs saved as Markdown and JSON (human-readable, version-control-friendly).
• SQLite persistence — no hidden databases, no external servers.
• Configurable extraction interval (default: every 5 minutes).
• Cross-platform — Windows 10+, macOS 12+, Linux (Ubuntu 22.04+ tested).
• PyInstaller distribution — single executable, no Python installation required for end users.

---

Privacy

Taskify is a privacy-first application. The following statements are guarantees, not marketing language:

• No network requests are made after initial model downloads. The application contains no telemetry, analytics, crash reporting, or update checks that contact external servers.
• All speech audio is processed in memory on your local CPU. No audio is written to disk unless you explicitly enable debug logging.
• All transcripts and extracted tasks are stored exclusively in your local SQLite database and log files, located at paths you control (see Configuration).
• The Ollama LLM integration communicates only with a process running on localhost. No cloud API key is required or accepted.
• The source code is fully auditable. You are encouraged to review src/taskify/audio/, src/taskify/stt/, and src/taskify/llm/ to verify these claims.

Data locations:

Data Type	Default Location
Database	~/.local/share/taskify/taskify.db
Daily transcripts	~/.local/share/taskify/logs/YYYY-MM-DD.md and .json
Application log	~/.local/share/taskify/app.log
Downloaded models	~/.local/share/taskify/models/
User configuration	~/.config/taskify/config.toml

To delete all application data, remove the directories listed above.

---

Requirements

Runtime:

• Python 3.11 or higher
• A working microphone
• 4 GB RAM minimum (8 GB recommended when running a local LLM)
• Approximately 1–4 GB free disk space for models, depending on selections

Optional (for LLM-based task extraction):

• Ollama installed and running locally
• One of the supported models pulled: phi4-mini, gemma3:2b, qwen2.5:1.5b

---

Installation

From Source (Recommended for Development)

# Clone the repository
git clone https://github.com/MrSpideyNihal/Taskify.git
cd Taskify

# Create and activate a virtual environment
python -m venv .venv

# Windows
.venv\Scripts\activate

# macOS / Linux
source .venv/bin/activate

# Install the package with all dependencies
pip install -e ".[dev]"

Minimal Install (Core Only, No LLM)

pip install -e ".[nlp]"

From a Pre-built Executable

Pre-built executables are available on the Releases page for Windows, macOS, and Linux. Download the archive for your platform, extract it, and run the taskify executable. No Python installation is required.

---

Model Downloads

Taskify does not bundle speech or language models. They are downloaded once on first run (or manually) and stored locally.

Vosk Speech-to-Text Models

Vosk models are available from alphacephei.com/vosk/models. The application will prompt you to download a model on first launch if none is found.

To download manually:

# English (small, ~50 MB) — recommended for most users
python -m taskify download-model --engine vosk --lang en --size small

# English (large, ~1.8 GB) — higher accuracy
python -m taskify download-model --engine vosk --lang en --size large

Supported language codes follow the Vosk model naming convention (en, de, fr, es, zh, hi, and others). See docs/model_download.md for the full list.

faster-whisper Models

python -m taskify download-model --engine whisper --model tiny
python -m taskify download-model --engine whisper --model base
python -m taskify download-model --engine whisper --model small

Ollama Language Models (Optional)

Install Ollama and pull a supported model:

ollama pull phi4-mini
# or
ollama pull gemma3:2b
# or
ollama pull qwen2.5:1.5b

If Ollama is not running, Taskify falls back automatically to the NLP backend.

---

Usage

Launch the GUI

taskify
# or
python -m taskify

Quick Start

1. Launch Taskify.
2. Select your microphone in Settings if the default device is incorrect.
3. Click Start Recording.
4. Speak naturally. The live transcript preview updates in real time.
5. Click Stop Recording when finished.
6. Task extraction runs automatically (or immediately via the Extract Now button).
7. Extracted tasks appear in the Eisenhower Matrix dashboard.

Command-Line Interface

# Run in CLI-only mode (no GUI, outputs tasks to stdout)
taskify --no-gui --duration 60

# Process an existing audio file
taskify process-file path/to/recording.wav

# Export tasks to Markdown
taskify export --format markdown --output tasks.md

# Show current configuration
taskify config show

---

Configuration

Copy the default configuration to your user config directory:

python -m taskify config init

This creates ~/.config/taskify/config.toml. All available options with their defaults:

[audio]
device = "default"          # Input device name or index
sample_rate = 16000         # Hz — Vosk and Whisper expect 16000
chunk_duration_ms = 200     # Audio chunk size fed to STT
silence_threshold = 0.01    # RMS energy threshold for silence detection
silence_duration_s = 1.5    # Seconds of silence before pausing capture

[stt]
engine = "vosk"             # "vosk" or "whisper"
vosk_model = "en-small"     # Vosk model identifier
whisper_model = "base"      # Whisper model size: tiny, base, small
language = "en"             # BCP-47 language code

[llm]
backend = "ollama"          # "ollama" or "nlp"
ollama_host = "http://localhost:11434"
ollama_model = "phi4-mini"
extraction_interval_s = 300 # Seconds between automatic extraction runs

[storage]
data_dir = ""               # Defaults to ~/.local/share/taskify
log_dir = ""                # Defaults to data_dir/logs

[logging]
level = "INFO"              # DEBUG, INFO, WARNING, ERROR
debug_llm = false           # Log full prompts and responses (verbose)

---

Architecture Overview

taskify/
  audio/          Audio capture, silence detection, chunk queue
  stt/            STT engine abstraction, Vosk backend, Whisper backend
  llm/            LLM abstraction, Ollama backend, NLP/rule-based backend
  pipeline/       Scheduler, segment processor, event bus
  storage/        DatabaseManager, transcript writer, migration runner
  export/         Markdown and JSON exporters
  ui/             Main window, matrix dashboard, settings dialog, widgets
  config/         Settings loader and validator
  cli/            Click-based CLI entry points

The core modules (audio, stt, llm, pipeline, storage) are intentionally free of any GUI dependency, enabling future headless and mobile deployments.

Data flow:

Microphone -> AudioCapture (sounddevice)
           -> STT Engine (Vosk / Whisper)  [daemon thread]
           -> TranscriptSegment queue
           -> TranscriptWriter             [background thread]
           -> SQLite + daily log files
           -> Scheduler (every N minutes)
           -> LLM / NLP Backend
           -> TaskItem list
           -> DatabaseManager
           -> EventBus -> UI refresh

---

Contributing

Contributions are welcome. Please read CONTRIBUTING.md before opening a pull request.

Development setup:

pip install -e ".[dev]"
pre-commit install

Running tests:

pytest --cov=src/taskify --cov-report=term-missing

Linting and type checking:

ruff check src/ tests/
mypy src/

Commit style: Conventional Commits (feat:, fix:, docs:, test:, chore:). No emojis in commit messages, code comments, or documentation.

---

Roadmap

See the GitHub Issues for the full backlog. High-level milestones:

Milestone	Target	Status
v0.1 — Core pipeline (audio, STT, storage)	TBD	In progress
v0.2 — LLM extraction and matrix dashboard	TBD	Planned
v0.3 — Settings, export, and packaging	TBD	Planned
v1.0 — Stable release with full documentation	TBD	Planned
v1.x — Android prototype via Chaquopy/Kivy	TBD	Research

---

License

MIT License. See LICENSE for full text.

Copyright (c) 2026 Nihal