FluidVoice Review: I Used the Viral Free Dictation App for 3 Weeks

For the past month, my feeds have been full of people raving about FluidVoice. Free, open source, on-device, with AI cleanup built in. It picked up 9,400 GitHub stars in under a year and a #1 ranking in an independent 21-app test. I write about dictation apps for a living, so I did the only thing that settles hype: I made it my daily driver for three weeks. Every email, every draft, every Slack message. Two app updates, one rollback, and a few mornings spent dictating into silence.

TL;DR: FluidVoice is the most feature-rich free dictation app on Mac. I score it 7/10 (3.5/5). It transcribes on-device, previews your words live as you speak, and cleans up your text with a local AI model — all at $0. But three weeks of real use surfaced the catch. My microphone setup broke three different ways. One update cut off my opening words until I rolled it back. The AI layer summarized what I said, or refused it like a chatbot. Analytics were on by default, and the AI runtime is closed-source, maintained by one person. My honest read: for tinkerers, FluidVoice is the best free dictation app on Apple Silicon today. If you dictate for work — or use an Intel Mac, macOS 13/14, or Windows — Voibe (the app we build) is $149 lifetime, pairs on-device Whisper with a zero-retention private cloud, and someone answers when it breaks.

FluidVoice is a free, open-source dictation app for macOS built in Swift by a solo developer under the ALTIC name. The ambition is right there in the repo's tagline: "Fastest and only macOS Dictation app with on-device STT and custom trained AI enhancement model. A local Wispr Flow alternative." No hedging — this is an app that wants to kill a $15/month subscription.

The growth curve backs up the hype. Created in September 2025, the project crossed 5,000 GitHub stars in late June 2026 and roughly doubled to about 9,400 stars and 620 forks by August 7, 2026 — I checked the numbers myself the morning I wrote this. The official site claims 100,000+ downloads (a first-party figure I couldn't verify independently). The license changed from Apache 2.0 to GPL v3 in February 2026.

The core loop is standard push-to-talk dictation: press a global hotkey, speak, and the transcribed text lands in whatever app is active. What sets it apart is the stack around that loop — a live word-by-word preview overlay while you speak, a selection of on-device speech models, and "Fluid Intelligence," a local AI cleanup layer with its own custom-trained model (Fluid-1). The requirements surprised me for a free app: macOS 15 (Sequoia) or later, Apple Silicon for all models (Intel Macs get the slower Whisper-only path since v1.5.1), about 1 GB of disk for a voice model, and roughly 3.5 GB more for Fluid Intelligence.

Install is brew install --cask fluidvoice or a DMG from GitHub releases. A Windows beta and an iOS waitlist exist; Linux is "planned." For where FluidVoice sits among free tools generally, see my best free dictation apps guide and our roundup of open-source Wispr Flow alternatives.

My setup for this test: an Apple Silicon MacBook Pro on macOS Sequoia, dictating with the built-in mic at my desk, AirPods on calls and walks, and a USB mic I plug in for recording days. Installation was a single brew install --cask fluidvoice, then a ~500 MB download for the default Parakeet model and another ~3.5 GB when I enabled Fluid Intelligence. Within ten minutes I was dictating, and I understood the virality immediately — this is a paid app's spec sheet at $0:

• Selectable on-device speech models: Nemotron Speech 3.5 (~670 MB, ~40 languages), Parakeet Flash (~250 MB, English, lowest latency), Parakeet TDT v3 (~500 MB, 25 languages, the default) and v2, Cohere Transcribe (~1.4 GB, 14 languages), Apple Speech, and Whisper Tiny through Large (~75 MB–2.9 GB, 99 languages). I stayed on Parakeet TDT v3 for most of the test.
• Live word-by-word preview: A transcription overlay shows your words as you speak — before they're inserted. This was the feature that hooked me in the first hour, and it's still rare in this category. (Voibe's Live Dictation offers the same on-screen-as-you-speak behavior.)
• Fluid Intelligence (Fluid-1): A local AI enhancement model (~3.5 GB) that removes fillers and reformats the transcript with no cloud, no API keys. Cloud providers (OpenAI, Groq, custom) are optional.
• Command Mode: Voice control of your Mac — open apps, run actions.
• Edit Mode and Write Mode: Select text and rewrite it by voice — though Edit Mode never actually worked for me (more below).
• Per-app configuration and custom dictionary: Different behavior per target app; user-defined vocabulary.
• Audio history: Recordings stored locally with an on-device transcript log.

Two honest caveats before you take the spec sheet at face value. The AI enhancement layer — the headline — turned out to be the least predictable part of my daily experience (I'll get to that). And the speed claims — "<100ms" perceived latency — are the developer's own numbers, not independent benchmarks. What I can say from use: Parakeet-based transcription on Apple Silicon felt genuinely fast, with text landing more or less as I finished speaking, and the v1.6.7 release notes claim a further 2x Parakeet speedup.

FluidVoice costs $0. No tiers, no word limits, no trial clock. On price alone, nothing beats it. Here's the three-year math against the category:

Using FluidVoice instead of Wispr Flow saves $432 over three years; instead of Superwhisper's annual plan, $254.97. Free wins the spreadsheet every time.

But three asterisks kept nagging at me as the weeks went on. First, the money model: the project runs on GitHub Sponsors, and the README states the Fluid Intelligence runtime is kept closed-source "so we can sustainably offer the core dictation experience for free" — which reads to me like a future monetization hook. That's fair enough, but it means the headline feature isn't actually the open-source part. Second, the support model: there is no support to buy at any price — when my microphone went silent, there was no one to email; bugs get fixed when the maintainer fixes them. Third, the time cost: I spent real hours of my test rolling back an update, re-selecting microphones, and re-dictating paragraphs that had landed in silence. Free software isn't free if it costs you a workday. For how I think about that trade generally, see our cloud vs local dictation breakdown.

This is the section that decides whether FluidVoice belongs in your daily workflow, so let me just walk through my test log. Week one was close to flawless — I genuinely started drafting a rave. Then the problems arrived in three distinct clusters.

1. Microphone handling — my biggest daily frustration

The first failure was the AirPods switch. I'd be dictating fine on the built-in mic, pop in AirPods for a call, come back — and dictation was dead. Not erroring, just silent: the overlay would appear, I'd talk, nothing. Twice I had to force-quit the app to get it listening again. Then, after updating to 1.6.7, FluidVoice quietly decided to use a different microphone than the one I'd selected — I dictated most of a paragraph into a dead input before I noticed nothing was appearing. And my USB mic never survived a round trip: I'd pin it as my input, unplug it at the end of a recording day (the app correctly fell back to the built-in mic), plug it back in the next morning — and FluidVoice would stay on the built-in mic until I manually re-selected the USB one. Every single time. If your setup involves AirPods, an external mic, or a docked laptop, expect this cluster to be your daily reality too.

2. The update that made things worse

FluidVoice ships updates roughly weekly, and the pace cuts both ways — two updates landed during my three weeks. One of them (1.6.6) introduced a noticeable delay after pressing the trigger key, and worse, started cutting off the first words of my dictations. "Can you send me the deck" would land as "send me the deck." For a dictation app, eating the start of my sentences is close to disqualifying. The app ships a "Get previous Builds" rollback button — and the fact that a rollback button exists as a first-class UI element tells you the developer knows updates break things. I used it, the problem went away, and I spent the rest of the test wincing slightly every time the update badge appeared.

3. The AI layer doing too much — or nothing, or something else entirely

Fluid Intelligence — the feature that makes FluidVoice more than raw transcription — was the least predictable part of my experience. What I want from cleanup is simple: basically verbatim what I said, minus the filler words. What I got, often, was a summary — I'd dictate five sentences of rambling context for an email and get back two tightened ones with my hedges and qualifiers gone. That's impressive technology and exactly wrong for dictation: those were my words, and I meant them. Some days enhancement silently didn't run at all and raw transcript came through, fillers and all — I never figured out the pattern. It also developed an obsession with formatting: after I enabled the editing features, every time I said the word "mention" it came out as "@mention," even in sentences that had nothing to do with tagging anyone. And once — my favorite failure of the whole test — I dictated a short phrase and my document received "I'm sorry, I can't assist with that." The cleanup model had decided my sentence was a prompt and refused it, like a chatbot. This is the known failure mode of running a content-rewriting LLM over dictation, and FluidVoice hasn't solved it. (Edit Mode, for what it's worth, never worked for me at all — I'd select text in VS Code or TextEdit, speak my change, and the app would fail to find the selection and just type over it in write mode instead.)

Fairness requires the other side, and I want to be clear about it: fixes ship fast — 1.6.7 fixed several mic problems even as it introduced my wrong-mic surprise — the maintainer is visibly responsive, and judging by the project's GitHub, plenty of users are having a better time than I did (and some a worse one). This is what a talented solo developer absorbing a viral influx looks like. But the pattern is structural: one person, weekly releases, no QA team, and a rollback button as the safety net. My testing across this category keeps finding the same thing: dictation daemons breaking on sleep, wake, and device changes is the #1 failure class for every dictation app — the difference is who's accountable for fixing it, and on what timeline.

FluidVoice's core privacy architecture is genuinely good — I want to say that plainly before the footnotes. Speech-to-text runs on-device, the local Fluid-1 enhancement model needs no API keys, and cloud AI providers are strictly opt-in. The README promises your voice and transcribed text "never leave your machine unless you explicitly opt in to a cloud AI provider." I dictated client names and unreleased product details into it for three weeks on that basis. For why that architecture matters, see why offline dictation matters.

Poking around settings and the project's documentation, though, three things jumped out at me:

• Analytics are on by default. A few days into my test I went through every settings pane and found anonymous analytics quietly enabled — I had to switch them off myself. The developer discloses that analytics exclude voice, audio, transcripts, and prompts, and the opt-out works. But opt-out-by-default is a choice worth knowing about in an app marketed as "zero data leaves your Mac," and it only got a toggle after users pushed back on the telemetry earlier this year.
• The headline AI is closed-source. Fluid Intelligence is "a private, separately maintained runtime" per the README. The app is open source; the feature that makes it special is not. You cannot audit the ~3.5 GB model runtime the way you can audit the Swift app around it.
• Bugs can become privacy bugs. An earlier build had a now-fixed bug where the dictation UI could surface transcriptions from unrelated sessions while idle. Fixed is good — but it illustrates that "local" and "leak-proof" are different properties, and a solo project has no security review process between you and the next such bug.

One more trust note for Windows users eyeing the beta: the unsigned Windows installer has been flagged by 3 of 70 VirusTotal vendors — almost certainly a false positive from missing code-signing, but exactly the kind of thing a signed, notarized commercial app doesn't ask you to shrug off.

What I Loved

• Free forever: No tiers, limits, or subscriptions — $0 against $144/year for Wispr Flow.
• On-device model buffet: Nemotron, Parakeet Flash/TDT, Whisper, Apple Speech, Cohere Transcribe — pick your speed/accuracy/language trade-off.
• Live word-by-word preview: Seeing my words appear as I spoke, before insertion, is the feature I missed most after the test ended. Rare and genuinely useful.
• Local AI cleanup with no API keys: Fluid-1 runs entirely on-device — no cloud round-trip for formatting.
• Independent validation: Top macOS pick in Adam Jones's 21-app Wispr Flow alternatives test — "the closest to Wispr Flow's UX of anything I tried."
• Fast-moving development: Two releases shipped during my three-week test; the Parakeet path got 2x faster in v1.6.7 alone.
• Power features at $0: Command Mode, Edit Mode, per-app configuration, custom dictionary.

What Wore Me Down

• Reliability under real use: A freeze when I switched to AirPods, the wrong mic silently selected after an update, a USB mic that lost its pinned status on every replug, and two force-quits to recover.
• Update roulette: One of the two updates during my test cut off my opening words until I used the built-in rollback button — which exists precisely because this keeps happening.
• AI layer unpredictability: Summarized my dictation, silently failed to run some days, turned "mention" into "@mention" everywhere, and once pasted a chatbot-style refusal into my document.
• Closed-source core feature: The Fluid Intelligence runtime can't be audited despite the open-source label.
• Default-on analytics in an app marketed on "zero data leaves your Mac" — I had to find the opt-out myself.
• Steep requirements: macOS 15+ only; Apple Silicon effectively required; ~4.5 GB disk with enhancement enabled.
• Solo-maintainer risk: Responsive today, but one person is the entire support, QA, and security team.
• Silently dropped short utterances (under ~1 second) — my quick "yes" and "thanks" replies kept simply vanishing.

After three weeks of daily use, here's my honest sorting of who should and shouldn't install it.

Best For

• Tinkerers on recent Apple Silicon Macs: If you run macOS 15+, enjoy trying models, and don't mind rolling back a bad build, FluidVoice is the most capable free dictation app you can install today.
• Wispr Flow subscribers who want off the subscription: It replicates more of Wispr Flow's UX (live preview, AI cleanup) than any other free tool, per Adam Jones's 21-app test.
• Budget-zero users who still want AI cleanup: Handy is free but outputs raw text; FluidVoice is the only free option with a local formatting layer built in.
• Open-source supporters who accept a pragmatic open-core arrangement and want to sponsor an indie developer.

Not Ideal For

• Anyone who dictates for a living: Writers, doctors, lawyers, and support teams need every dictation to land. Based on my three weeks — two microphone failures, one regression, one rollback — I can't call FluidVoice a safe daily dependency yet.
• Intel Mac and macOS 13/14 users: FluidVoice requires macOS 15+ and effectively Apple Silicon. Voibe supports macOS 13+ and Intel Macs through its zero-retention private cloud.
• Windows users: The beta is version 0.0.7 with an unsigned installer. Voibe for Windows is a shipped native app.
• People who need verbatim output: If AI enhancement summarizing or rewriting your words is unacceptable (contracts, medical notes, quotes), you want a bounded formatter — not a rewriter — or raw transcription.
• Anyone who needs support with an SLA: There is a Discord and a responsive maintainer, and that's the entire support organization.

I went looking for the usual third-party ratings and came up surprisingly empty: FluidVoice has no Product Hunt listing, no G2 profile, and no Trustpilot page as of August 2026 — unusual for an app this popular, and it means there is no aggregate rating to cite. What I could verify:

• GitHub: roughly 9,400 stars and 620 forks (August 7, 2026), up from ~5,000 in late June 2026 — a genuine viral trajectory.
• Adam Jones (independent, tested 21 Wispr Flow alternatives): named FluidVoice his top pick for macOS — praising the live preview and default-on cleanup — while flagging that "short utterances under 1 second get silently dropped" and "the default cleanup prompt is mediocre."
• Hacker News: word-of-mouth is positive in other apps' threads ("FluidVoice is incredible and developing quickly" — March 2026), but not unanimous: "I tried FluidVoice (found it buggy) and Handy (had problems with hotkey) before landing on TypeWhisper" (June 2026). FluidVoice's own HN submission received 2 points — its virality came from Reddit, X, and YouTube instead.
• Editorial roundups are split: #1 of 21 at adamjones.me, but mid-pack or last in other roundups (macaiapps ranked it #9 of 9; agenticcodingweekly #5 of 5). Coverage is genuinely mixed by outlet — worth knowing when a single glowing ranking crosses your feed.

The pattern across every source matched my own three weeks almost exactly: people love what FluidVoice does and are split on whether it keeps doing it reliably.

FluidVoice occupies a specific slot: free, on-device, feature-rich, fast-moving, fragile. Depending on which of those you'd trade, here's where I'd point you:

For dictation you can bet your workday on: Voibe

Voibe — the app we build — is my answer to everything in this review's cons list. It runs on Mac and Windows: the fully on-device Whisper mode is Mac-only (Apple Silicon), while Intel Macs and Windows use Voibe's zero-retention private cloud — self-hosted open-source models where audio is never stored, sold, or used to train AI. That architecture covers the exact hardware FluidVoice leaves out (Intel Macs, macOS 13/14, Windows) without giving up the privacy posture. Where FluidVoice's AI enhancement kept summarizing my dictation, Voibe's Smart Formatting is deliberately a bounded formatter: it removes fillers, fixes punctuation, and converts numbers — it does not paraphrase, change meaning, or generate content, and it's off by default. Add Live Dictation (the same words-as-you-speak preview FluidVoice is praised for), Developer Mode with VS Code, Cursor, and Windsurf file-name resolution, and a team accountable for stability rather than a rollback button. $7.50/month, $59/year, or $149 lifetime — a one-time price for the reliability that free doesn't include. Try Voibe for free.

For free with maximum simplicity and Linux: Handy

Handy is MIT-licensed, fully open source (no closed components), and runs on Mac, Windows, and Linux. It's simpler than FluidVoice — raw transcription, minimal punctuation, no AI layer — which also means fewer moving parts to break. Read my colleague's Handy review.

For open source with a support option: VoiceInk

VoiceInk is GPL v3 open source on Mac with a $29 one-time license that funds updates and support — a middle path between FluidVoice's free-and-fragile and a full commercial product. Read our VoiceInk review.

For maximum configurability on-device: Superwhisper

Superwhisper runs Whisper and Parakeet locally on Mac with deep per-app custom modes. $8.49/month or $249 lifetime; its common criticisms are setup complexity and price. Read our Superwhisper review.

For cloud polish and mobile: Wispr Flow

Wispr Flow — the app FluidVoice defines itself against — is the most polished cross-platform option (Mac, Windows, iOS, Android) at $15/month or $144/year, with cloud processing and a documented 2.7/5 Trustpilot rating. Read our Wispr Flow review, or browse every head-to-head on our comparison hub.

Yes — as a free experiment on the right hardware. Not yet — as infrastructure you depend on.

FluidVoice earns its 7/10 and its 9,400 stars. A free app with selectable on-device models, a live transcription preview, and local AI cleanup is a genuine achievement, and the developer's pace is impressive. Would I keep it on a spare Apple Silicon Mac to play with? In a heartbeat — it costs nothing to try and might be all you need.

Choose FluidVoice if you:

• Want the most capable free dictation app on Apple Silicon and accept rough edges
• Are leaving a Wispr Flow subscription and want the closest free equivalent ($432 saved over three years)
• Enjoy model-tinkering, Discord communities, and occasionally rolling back a release

Skip FluidVoice if you:

• Dictate professionally and can't absorb dead microphones, dropped words, or an AI layer that rewrites what you said
• Run an Intel Mac, macOS 13/14, or Windows
• Need verbatim, meaning-preserving output for legal, medical, or client work
• Want someone accountable when it breaks

Here's how I'd frame the decision after living with it for three weeks: on its good days, FluidVoice was the best free dictation experience I've had on a Mac, and the live preview spoiled me for other apps. On its bad days, I dictated into silence, lost the front of my sentences, and got a chatbot refusal pasted into a client email draft. That's a fine trade at $0 — until a lost dictation costs you more than a license. A brief due at 5 PM, a patient note, a client email — that's when the math flips, and when Voibe at $149 lifetime makes sense: the same on-device privacy on Apple Silicon, a zero-retention private cloud for every other machine you own, formatting that never paraphrases you, and a team whose job is making sure this morning's update didn't break your microphone.

Overall rating: 7/10 — a remarkable free tool I'll keep watching, with a reliability story you should hear before you rely on it.

• 6 Best Free Dictation Apps for Mac — where free tools shine and where they cost you
• Best Open-Source Wispr Flow Alternatives — the full open-source field compared
• Handy Review: Free Open-Source Offline Dictation — the simpler free alternative with Linux support
• Best Handy Alternatives — the free and paid field around the other viral open-source app
• VoiceInk Review: Open-Source Mac Dictation From $29 — open source with a support option
• Wispr Flow Review — the subscription app FluidVoice positions against
• Superwhisper Review — the on-device power-user option
• Best Offline Dictation Apps for Mac
• Why Offline Dictation Matters — the privacy architecture explained
• Cloud vs Local Dictation — the trade-off behind every app in this review
• Why Dictation Apps Stop Working After Sleep — the category-wide stability problem
• Voibe for Windows — native Windows dictation on a zero-retention cloud