Siri's offline cousin, built from spare parts

What it does

This is a Python script that listens to your voice, transcribes it with OpenAI’s Whisper, pipes the text to a Mistral 7B model running through Ollama, then speaks the response back using your Mac’s built-in text-to-speech. It is explicitly a Mac-compatible fork of maudoin’s original ollama-voice, with the README noting “various improvements” that are left unspecified.

The interesting bit

The entire pipeline runs offline — no API keys, no cloud latency, no subscription. The project leans entirely on Apple’s own speech synthesis and your local silicon, which makes it less a product and more a proof that modern open-source models have crossed the threshold where “voice assistant” no longer requires a data center.

Key highlights

• Fully offline operation: Whisper + Ollama + macOS TTS, no network required after setup
• Defaults to Mistral 7B and Whisper base.en, with YAML config for swapping models or languages
• Supports Apple Silicon, though PyAudio needs a Homebrew-installed portaudio shim
• Voice quality is upgradeable via macOS’s hidden “Premium” voices (the README suggests “Zoe”)
• 517 stars suggest a surprising appetite for voice UIs that don’t spy on your kitchen conversations

Caveats

• The “various improvements” over the original are not enumerated; you’ll need to diff the repos to find them
• Setup involves manual Whisper model placement and dependency wrangling that the README documents in nine steps
• No mention of wake words, interruption handling, or other niceties that make commercial assistants feel responsive

Verdict

Worth a Saturday if you’ve already got Ollama running and want to talk to your LLM without typing. Skip it if you need something that feels finished, or if you’re not on macOS — the Mac-specific dependencies are non-negotiable.