FluidVoice 1.6.1: The Open Source macOS Dictation App With On-Device STT and Fluid Intelligence

Paid dictation apps taught millions of Mac users that voice can be faster than typing — but most of them still route your speech through someone else's cloud.

FluidVoice takes the opposite default: open source dictation where transcription runs on your Mac, text inserts into any app via accessibility APIs, and optional enhancement can stay local too. The project crossed 5,000 GitHub stars ahead of v1.6.1 (released late June 2026), positioning itself as the fastest native Parakeet implementation on macOS and a credible local Wispr Flow alternative.

Install in one line:

brew install --cask fluidvoice

Or grab the latest build from GitHub Releases.

TL;DR

What's New in 1.6.x

Version 1.6.0 was the headline release; 1.6.1 shipped minor fixes three days later (version bump in Info.plist, stream-flag hardening in LLM request bodies per recent PRs).

The 1.6.0 changelog matters because it reframes what "fast dictation" feels like on Apple Silicon:

1. Insanely fast Parakeet — rebuilt Parakeet path with near-zero delay between speaking and seeing words on screen.
2. Fluid Intelligence — fully local AI model for on-device dictation enhancement: smart formatting, context-aware capitalization, post-processing. No cloud, no API keys.
3. Better theming — adaptive light/dark with a compact toolbar switcher.
4. Refreshed onboarding — language-first voice engine setup, a real dictation tryout, and optional Fluid Intelligence download in one pass.

The README's tongue-in-cheek warning is accurate: if local enhancement is good enough, you may cancel a paid dictation subscription.

Fluid Intelligence: Open App, Private Runtime

This distinction matters for open-source purists.

FluidVoice the app is GPLv3 (relicensed from Apache 2.0 effective 2026-02-23). You can clone, build, and audit the dictation shell, model integrations, overlay, and settings.

Fluid Intelligence is described as a separately maintained local AI runtime — not shipped as open source. It powers the premium on-device enhancement layer while the core dictation stack stays free. The team says keeping Fluid Intelligence private for now helps sustain free core development; that may change later.

Practically:

• Without Fluid Intelligence: FluidVoice still works with any supported speech model plus optional cloud LLM cleanup (OpenAI, Groq, custom).
• With Fluid Intelligence: Enhancement stays on your Mac — smart punctuation, capitalization, and post-processing without outbound API calls.

Neither path sends raw audio to FluidVoice's servers by default. Cloud providers only see text if you opt in and configure keys.

Speech Models: Pick Latency vs. Language Coverage

FluidVoice supports more on-device STT backends than most single-purpose dictation apps:

Parakeet TDT v3 covers Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Russian, Slovak, Slovenian, Spanish, Swedish, and Ukrainian.

Onboarding is language-first: pick coverage and latency, download once, dictate everywhere.

Core Features Beyond Raw Transcription

Live preview and overlay

Real-time transcription appears in a notch-aware overlay on MacBook Pros with a Dynamic Island-style notch, or a standard overlay elsewhere. Sizes run from minimal pill to large preview — all optional.

Global hotkey and smart typing

Hold your configured hotkey from any app. FluidVoice captures audio, transcribes locally, and inserts text through macOS accessibility APIs — the same integration pattern commercial dictation tools use, but under your control.

Command Mode

Voice-control your Mac: launch apps, run Shortcuts, trigger system actions, automate workflows. This is closer to agentic computer use than pure dictation — think "open Slack and message the team" rather than only pasting transcript text.

Write Mode

Dictate new content or rewrite selected text in place inside any text field — email clients, Notion, VS Code, terminal prompts. Select → speak → replaced or appended text.

History, stats, and per-app configs

Optional local audio history with budget controls and ZIP export. A today-usage stats header and toolbar pill track daily dictation volume. Per-app prompt sets let Slack dictation behave differently from code comments — fully optional.

Updates and beta channel

Sparkle-style auto-updates with an optional beta channel (Settings → Automatic Updates → Beta Releases) for early builds.

Privacy and Analytics

FluidVoice is local-first:

• Voice, raw audio, and transcribed text do not leave your machine unless you configure a cloud AI provider.
• API keys live in macOS Keychain; the app prompts for "Always allow" on access.

Anonymous analytics are on by default (app version, macOS version, low-cardinality feature flags, approximate usage ranges, success/error outcomes). Disable anytime: Settings → Share Anonymous Analytics. The README explicitly excludes voice, transcripts, window titles, file paths, clipboard, and typed content from collection.

Quick Start

1. Install: brew install --cask fluidvoice or download from releases.
2. Permissions: Allow microphone and accessibility — both are required.
3. Hotkey: Pick a global shortcut in Settings.
4. Onboarding: Choose a speech model for your language and latency target.
5. Optional — Fluid Intelligence: Download during onboarding for local enhancement (~3.5 GB).
6. Optional — cloud AI: Add OpenAI, Groq, or a custom endpoint if you want vendor LLM cleanup instead of or alongside Fluid Intelligence.
7. Optional — beta: Enable beta releases for early features.

Build from source:

git clone https://github.com/altic-dev/FluidVoice.git
cd FluidVoice
open Fluid.xcodeproj

Requires Xcode; dependencies resolve via Swift Package Manager. Unsigned CI build: xcodebuild -project Fluid.xcodeproj -scheme Fluid -destination 'platform=macOS' CODE_SIGNING_ALLOWED=NO.

How FluidVoice Compares to Other Mac Voice Tools

When FluidVoice fits best: You want dedicated dictation with model choice, minimal cloud dependency, and optional local enhancement — without running a full voice studio or a cloud agent that controls your entire desktop.

When Voicebox fits better: You also need voice cloning, TTS, and MCP so Claude speaks back in a cloned voice.

When HeyClicky fits better: You want always-on agentic Mac control powered by a cloud realtime model, not primarily transcription speed.

Community and Roadmap Hints

• Discord: discord.gg/VUPHaKSvYV
• Updates: @ALTIC_DEV on X
• Site: altic.dev/fluid
• Sponsors: GitHub Sponsors and Ko-fi linked from the README — funding message mentions future iOS and Windows ports

The repo's About line still reads: "Windows, iOS and Linux coming soon." Today shipping is macOS-only; Intel Macs lean on Whisper.

Related Reading

• Meetily: Privacy-First Local Meeting Transcription (21.3k stars) — Whisper/Parakeet meeting STT + Ollama summaries
• Aleph Neuro Silent Speech with Ultrasound (July 2026) — private voice input without audible speech
• Voicebox: The Free, Open Source AI Voice Studio That Replaces ElevenLabs and WisprFlow
• HeyClicky: Voice-Controlled Mac Demo Powered by GPT-Realtime 2.0
• Mercury 2 Diffusion LLM — Wispr Flow CTO on Real-Time Transcript Cleanup
• Cursor Big Day — Voice Dictation and Composer Pricing