Offline Voice Assistant

This repository contains an offline voice assistant implemented in Python. It uses Vosk for local speech-to-text, pyttsx3 for TTS, and optionally the Ollama CLI as a local LLM decider. The project provides a CLI and a lightweight Tkinter GUI.

Features

• Real-time local STT using Vosk (optional - falls back to typed input when missing)
• Local LLM decider integration via the ollama CLI (optional)
• TTS via pyttsx3 (optional)
• Safe execution mode (dry-run by default) and an option to allow direct execution of actions
• Clickable GUI launcher included: run_voice_assistant_ui.bat

Quick start (recommended)

1. Create a virtual environment and activate it (Windows PowerShell):

python -m venv .venv
.\.venv\Scripts\Activate.ps1

2. Install Python dependencies:

pip install -r requirements.txt

3. Make sure the Vosk model folder vosk-model-en-in-0.5/ is present in the project root. If not, download and place it in that folder.

4. Run the GUI:

python -m voice_assistant.gui
# or double-click run_voice_assistant_ui.bat

Notes about Ollama

• Ollama is used via its CLI binary (ollama). It's not a pip package. Install Ollama and ensure ollama is available on PATH or configure ollama_path in config.yaml.

Configuration

• Edit config.yaml to change defaults (wake word, allow_execution, use_ollama, etc.). Command-line flags are available via python -m voice_assistant.

Development

• The package is structured as a Python module voice_assistant/. Run python -m voice_assistant for CLI usage or python -m voice_assistant.gui for the UI.

Troubleshooting

• If the GUI shows "Simulated input - type or speak", Vosk or sounddevice may not be installed in the Python used to launch the GUI. Ensure you run the GUI using the same Python that has deps installed (the included run_voice_assistant_ui.bat tries local virtualenvs first).
• To see which Python the GUI uses, run it from a terminal and compare python -c "import sys; print(sys.executable)" output.

License

• This project is provided as-is. Add your preferred license if you plan to distribute. Offline Voice Assistant - skeleton

This repository contains a minimal, dependency-free skeleton for an offline voice assistant. It's intended as a starting point to integrate:

• Vosk (offline speech recognition)
• Ollama + local LLM (intent parsing)
• pyttsx3 (offline TTS)

What I created:

• voice_assistant/main.py - runner with run_interactive() and run_test() (one-shot smoke test)
• voice_assistant/stt.py - simulated STT (reads stdin)
• voice_assistant/nlp_model.py - tiny rule-based intent parser
• voice_assistant/executor.py - safe executor (dry-run by default)
• voice_assistant/config.yaml - wake word and options
• voice_assistant/memory.json - persistent memory
• voice_assistant/requirements.txt - recommended optional deps

Quick smoke test (one-shot, no external deps):

From the project root (parent of voice_assistant) run:

python -c "from voice_assistant.main import run_test; run_test()"

Or run the package directly (recommended) from the project root:

python -m voice_assistant --test

This will simulate: "hey gng open github" and print the parsed intent and dry-run result.

Next steps to integrate full features:

• Install Vosk and replace stt.listen() with a real audio capture + recognizer pipeline
• Integrate Ollama or another local LLM in nlp_model.parse_intent() for better intent/entity extraction
• Integrate Ollama or another local LLM in nlp_model.parse_intent() for better intent/entity extraction (now implemented; enable with use_ollama: true in config.yaml and install Ollama).
• Integrate Ollama or another local LLM in nlp_model.parse_intent() for better intent/entity extraction (now implemented; enable with use_ollama: true in config.yaml and install Ollama).
• LLM-driven decision mode: the LLM can also act as a decider that maps a user's utterance directly to an actionable JSON plan (open a URL, launch an app, tell the time). Enable with use_ollama_decider: true in config.yaml or pass --use-ollama --ollama-path <path> --debug-llm to the CLI. When enabled the LLM returns a JSON action object which the assistant validates and executes (or dry-runs if allow_execution: false).
• LLM-driven decision mode: the LLM can also act as a decider that maps a user's utterance directly to an actionable JSON plan (open a URL, launch an app, tell the time). Enable with use_ollama_decider: true in config.yaml or pass --use-ollama --ollama-path <path> --debug-llm to the CLI. When enabled the LLM returns a JSON action object which the assistant validates and executes (or dry-runs if allow_execution: false).

Wake-word / always-on listening

• By default this skeleton used a wake word (e.g., "hey gng"). If you find the wake word annoying you can disable it and let the assistant process all input immediately by setting always_listen: true in config.yaml or by passing --always-listen to the CLI. Use this with care - enabling always_listen means every captured phrase will be interpreted as a command.
• Enable allow_execution: true carefully in config.yaml when you're ready to allow system commands
• Add TTS with pyttsx3 in executor.execute() or a separate tts.py
• Add TTS with pyttsx3 in executor.execute() or a separate tts.py (now included). To enable voice output set allow_tts: true in config.yaml and install pyttsx3.