Best open-source dictation apps for Mac (2026)

Apple has shipped dictation on the Mac for over a decade, and most people who try it turn it off within a week. It transcribes what you said, literally, including every “um” and every false start, and then leaves you to clean it up. For a lot of writing, editing the transcript takes longer than typing would have.

So a whole category of replacements has grown up around it, and an unusual share of it is open source. That’s not an accident — dictation is a tool you leave running all day with access to your microphone and your keyboard, and a lot of people would rather read the source than trust a subscription.

Here are the ones worth knowing about in 2026, what each actually does, and the one distinction that matters more than any feature list.

The distinction that decides everything: local or cloud

Every app below falls on one side of a line, and picking the wrong side is the most common mistake people make.

Local apps run a speech model on your Mac. Audio never leaves the machine, they work on a plane, and they cost nothing to run. The tradeoff is that you’re limited to models small enough to run on your hardware, which generally means Whisper or Parakeet variants — good, but a generation or two behind what a hosted model can do on accents, proper nouns, and long alphanumeric strings. They also tend to be slower on older Macs, and they’ll spin your fans.

Cloud apps send audio to an API. You get larger, more accurate models and, increasingly, a cleanup pass that turns speech into finished text rather than a transcript. The tradeoff is real and worth stating plainly: your audio leaves your machine, you need a connection, and there’s a cost per hour.

Neither is correct in the abstract. If you dictate patient notes or privileged material, local is not a preference, it’s a requirement. If you dictate Slack messages and pull request descriptions all day and you’re tired of deleting your own filler words, cloud is going to feel meaningfully better.

Most of the honest disagreement in this category is just people weighting that tradeoff differently.

Blurt

Open source: MIT · Platform: macOS, Apple Silicon · Local or cloud: Cloud · Price: Free app, ~$0.62/hr of audio

Blurt is the clearest expression of the cloud side of the tradeoff, and it’s MIT-licensed, so you can read every line of it.

You hold the right ⌘ key, talk, and finished text lands in whatever app has focus. Not a transcript — finished text. That’s the part that’s different. Blurt sends your audio to the Dictation API, which runs transcription and an LLM cleanup pass in a single request, so self-corrections get resolved to what you actually landed on and filler disappears while your tone survives.

The worked example from the launch is a fair representation of what it does:

You say: “um so can we uh move the the meeting to thursday i think friday works better actually”

You get: “Can we move the meeting to Friday? That works better.”

Notice that it didn’t just strip the “um.” It resolved Thursday to Friday, because that’s what you meant by the end of the sentence. A transcript would have given you both.

It’s fast — around 134ms p50 on transcription, cleanup typically inside a second — and it handles 19 languages with mid-sentence code-switching, which matters more than it sounds if you’re one of the many people who switch languages mid-thought.

The honest downsides. It’s macOS-only and Apple Silicon-only, so there’s no Windows or Linux story at all. Audio goes to the cloud, full stop — if that’s disqualifying for your work, use Handy or FluidVoice instead and don’t think twice about it. And you need an AssemblyAI API key, which means it’s free to install but not free to run past your initial credits.

On privacy, the repo is specific rather than reassuring, which is the right way round. Your API key lives in the macOS Keychain. The request carries your recent dictations from the session and a short run of text before your cursor, to keep transcripts continuous — but never the app name, the window title, the field you’re in, or what you’ve selected, and it refuses to read password fields. History lives in memory only, capped and cleared on quit. No audio stored, no transcripts stored, no telemetry.

Handy

Open source: MIT · Platform: macOS (Intel and Apple Silicon), Windows, Linux · Local or cloud: Fully local · Price: Free

Handy is the purest expression of the local side, and it’s the one to recommend without hesitation if privacy is the deciding factor. Its README puts it simply: your voice stays on your computer.

It runs Whisper — Small through Large, plus Turbo — with GPU acceleration, and it also supports Parakeet V3, which is CPU-optimized and does automatic language detection. You can bring your own Whisper model if you have one. There’s a configurable global hotkey, push-to-talk, Silero VAD for detecting when you’re actually speaking, CLI parameters, and a Raycast integration.

It’s free, and deliberately so — the project’s stated position is that accessibility tooling doesn’t belong behind a paywall. Actively maintained, with a long commit history.

The tradeoffs. No cleanup pass — you get a transcript, and it’s on you to edit it. Larger Whisper models want real hardware, so an older Intel Mac will feel it. And the accuracy ceiling is whatever your local model can do, which on hard audio is below a current hosted model.

OpenWhispr

Open source: MIT · Platform: macOS, Windows, Linux · Local or cloud: Both · Price: Free

OpenWhispr has grown faster than anything else in this category over the past year, and the reason is that it refuses to pick a side on the local-versus-cloud question.

You can run Whisper and Parakeet models locally, or point it at a hosted provider with your own API key. Switch per task: local for anything sensitive, cloud when you want the accuracy.

The feature that deserves more attention than it gets is vocabulary learning — it picks up custom terms from your corrections over time rather than making you maintain a list by hand. There’s an AI cleanup step, a customizable global hotkey, and broad multilingual support with auto-detection.

The tradeoff is the flip side of the flexibility: it’s a lot of configuration surface for someone who just wants to talk into a text box, and the bring-your-own-key model means you’re managing credentials for whichever providers you enable.

FluidVoice

Open source: GPLv3 (earlier versions Apache 2.0) · Platform: macOS 15+ · Local or cloud: Local-first, optional cloud · Price: Free

FluidVoice is the most ambitious project here, and one of the most actively developed.

Core transcription runs on-device, across an unusually broad model list: Nemotron Speech 3.5, Parakeet Flash, Parakeet TDT v2 and v3, Cohere Transcribe, Apple Speech, and Whisper. Optional cloud enhancement through OpenAI, Groq, or a custom provider is available if you want it.

What sets it apart is that it isn’t only a dictation app. Command Mode lets you control the Mac by voice; Write Mode handles text entry. “Fluid Intelligence” does on-device formatting and capitalization. There’s a custom dictionary with spoken formatting commands, live preview with notch support, and per-app prompt configuration, so dictating into your terminal behaves differently from dictating into Mail.

The tradeoffs. macOS 15 minimum, so Sonoma users are out. It’s the most complex app on this list, which cuts both ways. And note the license change — it moved from Apache 2.0 to GPLv3 in February 2026, which matters if you were planning to build on it commercially.

TypeWhisper

Open source: GPLv3, commercial licensing available · Platform: macOS 14+, Windows, iOS · Local or cloud: Local by default, optional cloud · Price: Free for personal use; paid commercial licence

TypeWhisper is often described as fully local, which undersells what it does — its own framing is “local first, providers by choice.” Local engines are the default; cloud ASR and LLM providers come into play only once you configure and select them.

It ships with eleven local transcription engines, including WhisperKit and Parakeet, plus optional cloud providers. That breadth is the point: you can match the engine to the language and the hardware rather than taking whatever one model gives you.

The feature set leans toward power users: custom dictionaries, snippets, workflows, a plugin ecosystem, a REST API, and a CLI. It’s the most extensible option here.

The tradeoff is the licensing split — free personally, paid commercially — and the same complexity cost that comes with any tool this configurable. Check the current requirements and commercial pricing on typewhisper.com before committing.

VoiceInk (honorable mention — open source, not free)

Open source: GPLv3 · Platform: macOS 14.4+, Apple Silicon only · Local or cloud: Local transcription, optional cloud text enhancement · Price: One-time $25–$49

Worth including because the license is genuinely open even though the app isn’t free. VoiceInk runs Whisper via whisper.cpp and Parakeet via FluidAudio on-device, and any cloud enhancement is applied to the transcribed text rather than the audio.

Custom vocabulary, Smart Replace phrase expansion, filler-word removal, and automatic email and note formatting. Pricing is one-time — $25 for one device, $39 for two, $49 for three — with lifetime updates and no subscription. Its own repo positions it as the open-source alternative to Superwhisper and Wispr Flow, which is a fair description.

Side by side

How to actually choose

Four questions, in order:

Can your audio leave the machine? If no, you’re choosing between Handy, FluidVoice, TypeWhisper, and VoiceInk. Stop reading the rest of the list.

Do you want a transcript or finished text? This is the biggest practical difference in day-to-day use, and it’s the one people underestimate. A transcript is what you said. Finished text is what you meant. If you dictate a lot of messages and emails, the cleanup pass saves more time than any accuracy difference.

Are you Mac-only? Blurt and VoiceInk are. Handy, OpenWhispr, and TypeWhisper aren’t.

How much configuration do you want to own? OpenWhispr, FluidVoice, and TypeWhisper are deep, flexible, and require decisions. Blurt and Handy are close to zero-config in opposite directions.

The thing this category is quietly converging on

Two years ago every app on this list did the same job: run Whisper, paste the output. The differentiation was UI polish and hotkey handling.

That’s not true anymore. Look at what shipped in 2026 — OpenWhispr learning vocabulary from corrections, FluidVoice doing on-device formatting, TypeWhisper running a local LLM for cleanup, Blurt doing transcription and rewriting in a single API call. Every one of these is solving the same problem, which is that raw transcription was never the product. Nobody wants what they said. They want what they meant, formatted for where it’s going.

The local-versus-cloud argument will keep going, and it should. But the interesting question underneath it has already moved on: how much of the gap between speech and finished text can you close, and how much of that closing can you do without the audio leaving the room? The local projects are betting the answer is “most of it, eventually.” That’s a good bet to watch.

Frequently asked questions

What is the best free dictation app for Mac?

Handy is the strongest fully free option — it’s MIT-licensed, runs Whisper and Parakeet entirely on your Mac, and costs nothing to install or run. OpenWhispr is also free and adds the option of switching to cloud models when you want higher accuracy. Blurt is free to install but bills through an AssemblyAI API key at roughly $0.62 per hour of audio.

Is there a truly open-source alternative to Apple’s built-in Mac dictation?

Yes, several. Handy, OpenWhispr, and Blurt are MIT-licensed, while FluidVoice, TypeWhisper, and VoiceInk are GPLv3. All of them improve on macOS dictation in the same two ways: better speech models than Apple’s on-device engine, and some form of cleanup or formatting so you get usable text rather than a literal transcript.

Do open-source dictation apps work offline?

Handy works fully offline, and FluidVoice, TypeWhisper, and VoiceInk run their core transcription on-device. OpenWhispr works offline when you select a local model and requires a connection when you select a cloud one. Blurt requires a connection, since transcription and cleanup both happen server-side.

Which is better, Wispr Flow or Mac dictation?

Wispr Flow is substantially better than macOS dictation for most writing tasks, because it removes filler words, applies punctuation, and adapts tone per application, where macOS dictation returns a literal transcript. The tradeoffs are that Wispr Flow is a paid subscription beyond its 2,000-words-per-week free tier, and that its transcription always happens in the cloud. If open source matters to you, Blurt and VoiceInk cover similar ground with source you can read.

What’s the difference between a dictation app and a dictation API?

A dictation app is an end-user tool that puts text into whatever you’re typing in; a dictation API is what a developer calls to build that behavior into their own product. Blurt is an app; it’s built on the AssemblyAI Dictation API, which returns cleaned-up text and the verbatim transcript from a single HTTP request at $0.62 per hour. If you’re building dictation into your own software rather than looking for something to install, the API is the layer you want.

Do any of these support custom vocabulary for names and jargon?

OpenWhispr learns custom terms from your corrections over time, TypeWhisper and VoiceInk both support custom dictionaries, and FluidVoice has a custom dictionary with spoken formatting commands. Blurt supports key terms in Settings, which are passed with each request to bias spelling toward names and domain-specific words. This matters more than raw accuracy scores if your work involves a lot of proper nouns.