🧠 Kurtis-E1-MLX-Voice-Agent

A privacy-focused, offline voice assistant for macOS, powered by:

• 🧠 Local LLM inference via Ollama (replaceable with LM Studio or any OpenAI-compatible API endpoint)
• 🎤 Speech-to-text via MLX Whisper
• 🌍 Offline translations via Unbabel/TowerInstruct-7B-v0.2
• 🗣️ High-quality multilingual TTS (currently using XTTS v2)

This project is designed specifically for Apple Silicon Macs.

It prioritizes simplicity, speed, and on-device privacy for empathetic mental health conversations.

---

🛠️ Requirements

To run this project, you'll need:

• Python >=3.11
• Open-AI compatible API endpoint (Ollama, LM Studio, vLLM...)

Default Open-AI API endpoint (Ollama) is set as default. For LM Studio you can set the following environment variables:

# Remember to start LM Studio server.
export OPENAI_API_URL=http://localhost:1234/v1
export OPENAI_API_KEY=lmstudio

Remember to pull your model from Ollama (or LM Studio) before running the experiment.

# Default Ollama GGUF models

ollama pull hf.co/ethicalabs/Kurtis-E1-SmolLM2-1.7B-Instruct-Q4_K_M-GGUF
ollama pull hf.co/mradermacher/TowerInstruct-7B-v0.2-GGUF:Q4_K_S

🚀 Quick Usage

We recommend using uv as the Python runner:

uv run python3 -m kurtis_mlx

You can customize:

• --language: Select between english, italian, etc.
• --speaker: Change default speaker.
• --translate: Use your native language while chatting with an English-only LLM
• --llm-model: Defaults to Kurtis-E1 via Ollama
• --tts-model: Use a different voice model (e.g., XTTS v2)
• --whisper-model: Switch out Whisper variants

---

🔄 Goals

• ✅ Faster startup and playback (TTS runs in background worker)
• 🔐 100% offline: STT, LLMs and TTS run locally
• ☁️ Optional offline translation (only when --translate is enabled)