FluidVoice

Why it's ranked here

The core app is real open source, copyleft under GPLv3, but the feature the project promotes hardest, its private on-device enhancement layer, is closed and maintained outside the public repository. So privacy without inspection is the more accurate claim than privacy through openness. What is verifiable holds up well: several distinct speech engines spanning a fast light model to a slower, high accuracy multilingual one, cloud API keys handled through the system credential store instead of a config file, and a realtime audio pipeline built for one job. That combination, a transparent shell wrapped around a private add-on, is an unusual shape for this category, not a broken promise.

What's good

Cloud provider keys go into the macOS system credential vault as a single entry, with old per-provider entries swept up and removed automatically rather than left behind. Requests aimed at a local address skip attaching a cloud key entirely, so a locally hosted OpenAI-compatible server never receives a credential meant for someone else's API. The realtime audio capture path is a fixed-size, allocation-free ring buffer feeding a dedicated audio callback, the standard shape for code that cannot afford to stall while a person is mid-sentence. Several speech engines ship together, covering a real range from a fast small model to slower, more accurate multilingual options.

Tradeoffs

The most differentiated feature, the private local enhancement model, sits entirely outside the GPLv3 codebase and is described as staying that way for now, so the part doing the most interesting engineering cannot be read, audited or forked the way the surrounding app can. GPLv3 itself is a strong copyleft that limits how the app's own code can be reused elsewhere. None of the fetched files include a test suite; a plan for automated UI test coverage exists only as notes for a separate, unmerged branch. Every bundled speech model outside the Whisper family also needs Apple Silicon, so Intel Mac owners get a narrower set of choices.

How to use it well

It fits someone who dictates constantly across many different Mac apps and wants the audio and text handled locally by default, with a hotkey driven workflow instead of a separate recording step. Picking a smaller, faster speech engine suits quick notes; a larger multilingual one suits longer or non-English dictation. Anyone who wants the cleanup step to stay fully auditable should leave it off and either accept the raw transcript or route enhancement through their own cloud key, since the private local model cannot be inspected. It is a macOS-only tool today: iOS and Windows support is not shipped, only a waitlist signup.

Observed

License

GPLv3 for the application code

AI enhancement layer

The on-device enhancement model ('Fluid Intelligence') is a separate, privately maintained runtime distributed outside the GPLv3 codebase

Packaging and install surface

Homebrew cask, direct release download, or build from source with Swift Package Manager and Xcode

Platform support

macOS 15.0 or later; Apple Silicon required for the bundled speech and enhancement models, Intel Macs limited to the Whisper models

Interfaces

Desktop menu bar application; also starts a local API server on launch

Credential storage

Cloud provider API keys are stored in the macOS Keychain rather than in a plain settings file

Structural observation

No test directory or test target is present among the fetched repository files