What does "local" or "on-device" speech-to-text actually mean?

It means the audio is transcribed by a speech model running on your own Mac, and the recording never leaves the machine. There is no upload to a server, no account, and no internet round-trip in the transcription path. Cloud dictation does the opposite: it streams your audio to a remote service that runs the model and sends text back. The practical test is Airplane Mode. If dictation still works with the network off, the model is running locally.

Is local speech-to-text as accurate as the cloud?

On a modern Apple Silicon Mac, yes, and often more accurate. The model Parakeety runs, Parakeet TDT 0.6B v3, posts a 6.32% word error rate against Whisper Large V3’s 7.44% on the Hugging Face Open ASR Leaderboard. Accuracy is a property of the model, not of where it runs. A good local model on the Apple Neural Engine matches or beats most cloud dictation services, with none of the latency of a network round-trip.

Do I need an internet connection for local dictation?

Only once. Parakeety downloads the speech model on first launch, which is a one-time transfer of around 600 MB. After that, transcription runs entirely on-device and works offline. The only later network traffic is a periodic license check, never your audio.

What is the best local speech-to-text app for Mac?

It depends on the job. For push-to-talk dictation that pastes at the cursor in any app, Parakeety runs the Parakeet TDT v3 model on-device for $30 once. For transcribing pre-recorded audio files, a Whisper-based app like MacWhisper fits better. For hands-free computer control rather than dictation, Talon Voice is the category. The round-up of the trade-offs is in the best Mac dictation apps guide.

← Resources

The complete guide to local speech-to-text on Mac

Short answer: local speech-to-text means the model that turns your voice into text runs on your own Mac, so the audio never leaves the machine. On Apple Silicon this is now fast and accurate enough to replace cloud dictation outright. The shortest path to it is Parakeety, a push-to-talk Mac app that runs NVIDIA’s Parakeet TDT v3 model on the Apple Neural Engine for $30 once. This guide is the map of the whole category: what on-device transcription is, why it matters, which apps run it, and how to pick.

The landscape at a glance

Every Mac dictation tool sits somewhere on a line from fully cloud to fully on-device. Here is where the main options land, and what each one is actually built for.

Tool	Where it runs	Pricing model	Best for
Parakeety	On-device, always	$30 once	Push-to-talk dictation that pastes at the cursor in any app
Apple Dictation	On-device on modern Macs, with exceptions	Free, built into macOS	Short, casual dictation in mainstream languages
MacWhisper	Local transcription; AI features call the cloud	Free tier and paid Pro	Transcribing pre-recorded audio and video files
SuperWhisper	Hybrid: local and cloud models	Subscription or lifetime	A multi-engine menu with AI post-processing
Wispr Flow	Cloud	Subscription	Cloud dictation with AI cleanup, cross-platform
Dragon	Cloud (Dragon Medical One); no current Mac desktop app	Subscription	Enterprise medical dictation on Windows

Why local at all

The case for on-device transcription is three things at once: privacy, speed, and longevity. When the model runs on your Mac, the audio is captured to memory, transcribed, pasted at the cursor and discarded. Nothing is written to disk and nothing is uploaded. There is no data lake to subpoena, no breach surface, and no question of what a vendor logs.

The honest qualifier is that "local" is a spectrum, not a checkbox. Several popular apps transcribe locally but route the transcript through a cloud large language model for cleanup, which puts your words back on the network. So the question is rarely "is it local" but "which part is local". The detail per app is worth checking: whether SuperWhisper keeps audio on the device, whether MacWhisper’s AI features stay on the Mac, whether Apple Dictation runs on-device in your language, and the simpler offline test of whether each one works without an internet connection at all.

Speed is the part you feel most. A cloud service adds a network round-trip to every utterance; a local model on the Apple Neural Engine does not. And longevity matters for a tool you lean on daily: a subscription service can change its pricing, its model or its terms underneath you, or shut down. An app that runs on your own machine keeps working.

The model underneath

Accuracy is a property of the speech model, not of where it runs. The model that has changed what is possible on a Mac is NVIDIA’s Parakeet TDT 0.6B v3: roughly 600 million parameters, a transducer architecture, and the current top of the Hugging Face Open ASR Leaderboard. The full primer on what Parakeet is and how to run it on a Mac covers the technical detail.

The short version of why it matters: Parakeet TDT v3 posts a 6.32% word error rate against Whisper Large V3’s 7.44%, runs roughly an order of magnitude faster on the same hardware, and uses a transducer design that produces silence during silence instead of hallucinating text the way Whisper-based systems can. The trade is language coverage, 25 European languages against Whisper’s hundred or so. For dictation in those languages on Apple Silicon, it is the model to want.

Choosing an app

The right tool depends on the shape of the work. The honest 2026 round-up of the best Mac dictation apps walks through each option’s strengths and where it falls short. If you are weighing Parakeety against a specific competitor, the head-to-head pieces go deeper than a table can:

Parakeety vs Apple Dictation for the free baseline that ships with every Mac.
Parakeety vs SuperWhisper for the multi-engine, AI-post-processing comparison.
Parakeety vs MacWhisper for live dictation against file transcription.
Parakeety vs Wispr Flow for the on-device against cloud-subscription question.
Parakeety vs Dragon for anyone left without a desktop Dragon on Mac.
Parakeety vs Talon Voice for dictation against full hands-free control.

Who it is for

Local dictation earns its keep when the transcript is real work rather than a quick text message, and the privacy posture is part of the job. The audience primers cover the specific shapes of that: clinicians and GPs, therapists and counselors, lawyers and solicitors, writers and researchers, engineers, and translators working across languages.

Regulated work and compliance

For clinical and legal dictation the privacy question becomes a compliance question, and there are two valid answers. One is cloud dictation under a Business Associate Agreement; the other is on-device dictation where the audio never leaves the Mac in the first place. The difference between architectural and contractual privacy is the framing that decides which fits your practice. The specifics per tool are worth reading before you commit: whether Wispr Flow is HIPAA compliant, what Dragon Medical offers under a BAA, and where Apple Dictation lands for clinical use.

Dictating inside the apps you already use

Push-to-talk dictation that pastes at the cursor works anywhere a Mac accepts keyboard input, which means it slots into the tools you already live in. There are specific workflow guides for VS Code, Cursor and the JetBrains IDEs on the engineering side; Microsoft Word, Google Docs and Scrivener for writing; and clinical and legal systems like Epic, SimplePractice and Clio.

About this guide

I make Parakeety, a one-person Mac app. I write these guides from inside the problem: I built the app because the existing dictation options all asked for too much, so the comparisons here are honest about where Parakeety is the wrong choice as well as where it is the right one. If you want the background, I have written about why I built Parakeety and how I shipped a 600 MB speech model inside a 2 MB app.

FAQ

What does "local" or "on-device" speech-to-text actually mean?: It means the audio is transcribed by a speech model running on your own Mac, and the recording never leaves the machine. There is no upload to a server, no account, and no internet round-trip in the transcription path. Cloud dictation does the opposite: it streams your audio to a remote service that runs the model and sends text back. The practical test is Airplane Mode. If dictation still works with the network off, the model is running locally.
Is local speech-to-text as accurate as the cloud?: On a modern Apple Silicon Mac, yes, and often more accurate. The model Parakeety runs, Parakeet TDT 0.6B v3, posts a 6.32% word error rate against Whisper Large V3’s 7.44% on the Hugging Face Open ASR Leaderboard. Accuracy is a property of the model, not of where it runs. A good local model on the Apple Neural Engine matches or beats most cloud dictation services, with none of the latency of a network round-trip.
Do I need an internet connection for local dictation?: Only once. Parakeety downloads the speech model on first launch, which is a one-time transfer of around 600 MB. After that, transcription runs entirely on-device and works offline. The only later network traffic is a periodic license check, never your audio.
What is the best local speech-to-text app for Mac?: It depends on the job. For push-to-talk dictation that pastes at the cursor in any app, Parakeety runs the Parakeet TDT v3 model on-device for $30 once. For transcribing pre-recorded audio files, a Whisper-based app like MacWhisper fits better. For hands-free computer control rather than dictation, Talon Voice is the category. The round-up of the trade-offs is in the best Mac dictation apps guide.

Try it

Parakeety is a Mac menu-bar app. Hold the section key, talk, release; your words paste at the cursor in whichever app you were typing into. Audio never leaves the machine. There is a free 7-day trial with no card required. After that it is $30 once.

Try Parakeety free →