FluidVoice

Use multiple on-device speech models and a global hotkey for app-wide dictation

Apple Speech, Whisper, Parakeet, Nemotron, and other models can power live transcription, with text inserted through macOS accessibility APIs into arbitrary text fields. Apple Silicon gets the fastest model paths while Intel Macs remain supported through Whisper and Apple Speech.

Sources: [2]

Switch between local enhancement and optional cloud providers

The README says voice and text stay local unless a cloud AI provider is explicitly selected. OpenAI, Groq, and custom providers are optional, with keys stored in macOS Keychain. The advanced Fluid Intelligence local enhancement runtime is privately maintained separately from the core open-source app.

Sources: [2]

For people who want low-latency, privacy-oriented dictation across their entire Mac

It targets daily dictation in editors, browsers, chat, mail, and other text fields rather than only meeting transcription, with model choices based on language, latency, and hardware.

Account for microphone/accessibility permissions, model storage, analytics settings, and the private-runtime boundary

Dictation requires microphone and Accessibility permissions. Models range from hundreds of megabytes to several gigabytes, and Fluid Intelligence is about 3.5 GB. The README documents weekly activity signals and default-on detailed anonymous analytics that can be disabled. Fluid Intelligence itself is private, so it does not share the core app's GPL source availability.

Sources: [2][3]