FluidVoice is a macOS dictation app that transcribes speech on-device and types the result into whatever text field you are in. You hold a global hotkey, talk, and the text is inserted through the accessibility APIs, so it works the same in email, documents, chat, a terminal, or a code editor. A live preview overlay shows words as they appear, with a variant shaped around the MacBook notch.
Transcription runs against a speech model you choose during onboarding, trading language coverage against latency and download size. On top of that sits an optional enhancement layer that rewrites the raw transcript — stripping filler, fixing capitalisation, and formatting structure. That layer can run through Fluid Intelligence, a separate local AI runtime kept private by the project, or through a cloud provider whose key you supply and which is stored in the macOS Keychain.
Beyond plain dictation there are two other modes: Write Mode rewrites or replaces selected text in any app, and Command Mode drives the Mac by voice to launch apps, run shortcuts, and trigger system actions. Optional extras include local audio history with budget controls and ZIP export, daily usage stats, and per-app prompt sets so dictation adapts to the app you are in.
FluidVoice requires macOS 15.0 or later, plus microphone and accessibility permissions. Apple Silicon Macs can run every model; Intel Macs are supported through the Whisper models. Voice, audio, and transcribed text stay on the machine unless you opt in to a cloud enhancement provider; the app sends one anonymous activity signal per day, with more detailed anonymous analytics on by default and switchable off.
Features
- On-device transcription: local speech models convert audio to text without sending it to a server
- Speech model choice: Nemotron Speech 3.5, Parakeet Flash, Parakeet TDT v3 and v2, Cohere Transcribe, Apple Speech, and Whisper
- Local enhancement: Fluid Intelligence, a separate local AI runtime, handles formatting, capitalisation, and post-processing offline
- Cloud enhancement: optional post-processing through OpenAI, Groq, or a custom provider, with keys held in the macOS Keychain
- Write Mode: dictate new text or rewrite selected text in place in any app
- Command Mode: launch apps, run shortcuts, and trigger system actions by voice
- Live preview: a real-time transcription overlay, with a notch-aware variant and configurable size
- Global hotkey: capture starts from anywhere without switching apps
- Per-app configuration: assign different prompt sets to different applications
- Audio history: optional local recording history with budget controls and ZIP export
- Language coverage: roughly 40 languages on Nemotron, 25 on Parakeet TDT v3, 14 on Cohere Transcribe, and up to 99 on Whisper
- Intel support: Whisper models run on Intel Macs; the other models require Apple Silicon
