About this project
FluidVoice is an open-source voice-to-text dictation application for macOS, positioned as a local alternative to cloud dictation services. It is distributed under GPLv3 and installable via Homebrew cask or manual release download.
Core capability is speech-to-text dictation triggered by a global hotkey, with text inserted directly into any app through accessibility APIs. A live preview overlay shows transcription as you speak, including a notch-aware layout for MacBooks.
Speech recognition runs on-device. Supported models include Nemotron Speech 3.5 variants, Parakeet Flash, Parakeet TDT v3 and v2, Cohere Transcribe, Apple Speech, and Whisper sizes from Tiny to Large. Language coverage varies by model: Parakeet TDT v3 covers 25 languages, Cohere Transcribe 14, Nemotron roughly 40, and Whisper up to 99. Apple Silicon is required for most models; Intel Macs are supported through Whisper.
Beyond plain dictation, the app offers Command Mode for controlling the Mac by voice (launching apps, running shortcuts, triggering system actions) and Write Mode for writing or rewriting text in any text field. Optional AI enhancement can post-process transcripts using OpenAI, Groq, custom providers, or a bundled local model called Fluid Intelligence, which handles formatting, capitalization, and cleanup without sending data off the machine. Per-app prompt sets are configurable.
Other features include optional local audio history with budget controls and ZIP export, daily usage statistics, adaptive light/dark theming, menu bar integration, auto-updates with an optional beta channel, and configurable overlay sizes. The project states it is local-first: voice, audio, and transcribed text stay on the device unless a cloud AI provider is explicitly enabled. Anonymous weekly activity analytics are sent by default and can be disabled.
Requirements are macOS 15.0 or later, microphone access, accessibility permissions, and roughly 1 GB of disk space for a speech model (about 3.5 GB more for the optional Fluid Intelligence model). Building from source uses Xcode with Swift Package Manager; a build script supports signed and unsigned builds, and integration tests run via xcodebuild. The README notes iOS and Windows versions are planned but not yet available.
Comments
0 Rating appears after 10 ratings
Sign in to join the discussion.