VoiceInk: Open-Source Mac Dictation Explained

VoiceInk is a Mac dictation app that types what you say into any application ¹, using speech models that run on your own computer ². Its source code is public under the GNU GPL v3 ³, so you can build it yourself for free ⁴, or buy a one-time license: $25 for one Mac ⁵, $39 for two ⁶ or $49 for three ⁷. There is no subscription.

Three points decide whether it suits you. First, the default speech model, Parakeet V3, is a good choice for English: in our tests it made fewer errors than Whisper in background noise and on accented speech. Second, your audio stays on the Mac unless you pick a cloud model, but the optional AI enhancement sends your text, and any screen or clipboard context you enable, to whichever AI provider you choose. Third, it is Mac only, and needs an Apple silicon Mac for local models.

This guide covers what VoiceInk is, what the license allows, which model to pick, what its Modes and AI features send where, and how it compares with Superwhisper, Wispr Flow and the dictation built into macOS. For the whole field, see our overview of dictation software.

VoiceInk is open-source dictation that runs on the Mac

VoiceInk does one job: you press a shortcut, speak, and the text appears where your cursor is. Its README calls it a native macOS application that transcribes speech almost instantly ¹⁵. It is a form of system-wide dictation: it works in Mail, Slack, a code editor or a browser, not only in its own window.

Under the hood it is built on open components. The README credits whisper.cpp for running OpenAI's Whisper and FluidAudio for NVIDIA's Parakeet ¹⁶. Both run on the Mac itself, which is what on-device speech recognition means. The project's FAQ says all voice transcription is processed locally and that your voice data never leaves your Mac ¹⁷. That holds for the local models. It does not hold if you choose a cloud model, as the privacy section below explains.

The project is popular among developers. The GitHub repository had 6,935 stars on October 27, 2026 ¹⁴, and the name appeared in 21 Hacker News stories and comments in the 12 months to September 2026 ¹⁸.

The requirements differ between the maker's own pages

The pricing page says VoiceInk works only on Apple silicon Macs and needs macOS 14.4 or later ¹⁹. The installation guide says macOS 15 or later, and recommends cloud transcription on Intel Macs ²⁰. The model documentation says the app warns that local models do not work reliably on Intel Macs ²¹. In practice: on an Intel Mac, VoiceInk is not a local dictation app, so its main advantage disappears. On an M-series Mac with a current macOS, it works as described. The glossary entry on Apple silicon explains which Macs qualify.

The app has changed fast

VoiceInk added NVIDIA's Parakeet v2 and v3 models in version 1.60, released on October 29, 2025 ²⁴. Version 2.0, released on July 16, 2026, introduced Modes for workflows across apps, websites and tasks, added new AI providers and custom models, and introduced an Assistant ²⁵. On October 16, 2026 the newest stable tag on GitHub was v2.21 ²⁶. Reviews written before July 2026 describe an older app, so check their dates.

The iPhone and iPad app is a separate product. It is listed as free on the US App Store ⁸, and a Mac license does not activate it ²⁷. Its US rating was 4.26 stars ²⁸ from 38 ratings ²⁹ on October 26, 2026, and 4.59 stars from 22 ratings in Germany ³⁰. Those are small samples, and they rate the phone app, not the Mac app.

The flow below shows how one dictation moves through the app. Every box after the microphone can be local or remote, depending on what you choose.

The licence: build it free or buy it

VoiceInk's code is free to read, change and build, and the paid license buys convenience: a signed app, automatic updates, priority support and future features ⁹. The README states plainly that you can build it yourself from source, while the compiled version adds updates, support and funding for development ⁴. It also installs through Homebrew with brew install --cask voiceink ³¹, which fetches the same compiled app that needs a license or trial.

The GPL v3 is a copyleft license: anyone who distributes a modified copy must share its source under the same terms. The LICENSE file in the repository is only a short header that points to the full text online ³², which can confuse automatic license detectors, but the README is explicit. The project is not accepting pull requests; you may fork and modify it for your own use and report bugs through issues ³³. The guide to open-source dictation software compares VoiceInk with Handy, OpenWhispr and HEX, and the question whether VoiceInk is open source covers the licensing detail.

Three one-time plans, priced by number of Macs

Each is a perpetual license with updates included ⁹. You can try the app for 7 days first ¹⁰, and the pricing page offers a 14-day money-back guarantee, no questions asked ¹¹.

Two costs sit outside that price. The iPhone app is licensed separately ²⁷. And cloud AI is bring your own key: VoiceInk connects to cloud transcription and AI providers with your API key, and its license does not include their charges ³⁴. If you only use local models, there is nothing more to pay.

Over a year, VoiceInk costs a fraction of its subscription rivals

One year of Wispr Flow Pro billed yearly costs $144 ³⁵, about 5.8 times VoiceInk's Solo license ³⁶. Superwhisper Pro costs $84.99 a year ³⁷, or $249.99 once for a lifetime license ³⁸, about 5.1 times VoiceInk's largest plan ³⁹. The comparison is not like for like: Wispr Flow's price includes its cloud processing, and Superwhisper's license covers Windows and phones too ⁴⁰.

For what cloud AI adds to the bill when you bring your own key, see the guide to what AI dictation costs.

Parakeet or Whisper: what our measurements say

For most English dictation, keep the default, Parakeet V3. VoiceInk's own model guide names it the starting point for local dictation on Apple silicon and uses it during onboarding ¹³. Our measurements back that choice. Whisper Large v3 Turbo was slightly more accurate on clean read English, but Parakeet made fewer errors in background noise, on accented conversation and on one long recording, and it was faster. Pick Whisper when you need a language Parakeet does not cover.

We ran both models directly on a Mac with an M2 Pro in October 2026, outside VoiceInk, on public test sets. The measure is word error rate: the share of words wrong, missing or added, so lower is better. Apple's own on-device model, which VoiceInk offers as Apple Speech on macOS 26 ²², is included for reference. VoiceInk's README promises 99% accuracy ⁴¹ without saying under what conditions; the numbers below show how much the conditions matter.

In a quiet room the two are close, and Parakeet is faster

On 100 clean English recordings, Whisper Large v3 Turbo had a 3.92 percent error rate ⁴³ against 4.48 percent for Parakeet ⁴². You would rarely notice that gap. Parakeet transcribed the same audio about 19 times faster than real time ⁵⁷, against 12 times for Whisper ⁵⁸. In VoiceInk, Whisper also transcribes only after you stop speaking, rather than streaming partial text ⁵⁹.

In other European languages the picture is mixed. In German, Parakeet scored 4.84 percent ⁶⁰ and Whisper 5.08 ⁶¹. In French, Parakeet scored 5.29 ⁶² and Whisper 6.62 ⁶³. In Spanish, Whisper edged ahead at 2.53 ⁶⁴ against 2.79 ⁶⁵.

Noise and accents are where Parakeet pulls ahead

Background talk is the common real-world problem, and here the ranking flips. With babble at the level of a busy cafe, Parakeet stayed at 4.48 percent ⁴⁵ while Whisper rose to 5.51 ⁴⁶. In a loud cafe, Parakeet reached 7.85 ⁴⁸ and Whisper 13.92 percent ⁴⁹, many more errors. When the background voices were as loud as the speaker, no engine was usable: Parakeet hit 44.47 percent ⁶⁶ and Whisper 66.09 ⁶⁷. The guide to dictation in background noise covers microphones and settings that help.

On accented, conversational English from 13 accent groups, Parakeet had a 12.57 percent error rate ⁵¹ and Whisper 17.66 percent ⁵², a gap of 5.09 points, or about 40 percent more errors for Whisper ⁶⁸. If you speak English with an accent, or dictate the way you talk rather than reading aloud, that gap matters more than the quiet-room result. Our guide to speech recognition and accents has the results by accent group.

Long single recordings cost Whisper more

If you dictate for minutes without stopping, Whisper loses more. Sent as one long recording, Whisper's error rate rose by 0.56 percentage points against the same audio cut into pieces ⁶⁹, and it left out a passage of 10 or more words in 4 of 84 readings ⁷⁰. Parakeet rose by 0.25 points ⁷¹ and left out such a passage in 1 of 84 ⁷². Apple's model showed no measurable cost ⁷³. These tests ran each engine with its own long-audio handling, not VoiceInk's, so VoiceInk's results on long dictation could differ.

Language decides the rest

Parakeet V3 covers 25 European languages and detects which one you are speaking ⁷⁴. OpenAI says Whisper supports 98 languages ⁷⁵, with accuracy that varies by language. VoiceInk adds other local options: Cohere Transcribe for 14 languages including Arabic, Japanese and Korean ⁷⁶, and Nemotron Multilingual for languages in other scripts such as Chinese and Hindi ⁷⁷.

The questions on which VoiceInk model to use and which languages VoiceInk supports go further, and the guide to Whisper's languages shows how accuracy varies across them.

Modes and AI enhancement send text to a model you choose

Transcription in VoiceInk can stay entirely on your Mac, but AI enhancement is a second step that sends your words to whichever model you select. If that model is in the cloud, your text and any context you allowed leave the Mac. If it is VoiceInk Refine or a local Ollama server, they stay.

Modes switch settings by app, website or trigger word

Version 2.0 introduced Modes for per-app, per-website and per-task workflows ²⁵; the older name for this kind of switching was Power Mode, whose interface release 1.60 changed ²⁴. A Mode bundles a transcription model, a language, an optional AI step and an output style. VoiceInk picks a Mode in four ways: by the active app, by the website open in your browser, by a spoken trigger word, or by a dedicated shortcut ⁷⁸. Website detection works in Safari, Chrome, Arc, Edge and others, but not in Firefox ⁷⁹.

The starter Enhancement Mode cleans up dictated text with AI and uses selected text and screen context when they are available ⁸⁰. You can also select a passage and say how it should change, and VoiceInk replaces it in place ⁸¹. A personal dictionary keeps names and unusual spellings intact ⁸²; see custom vocabulary. For how this kind of AI dictation cleanup works in general, read the guide to AI dictation cleanup.

What leaves your Mac depends on two choices

Screen context deserves a careful look. When it is on, VoiceInk reads the visible text of your active window, which may be an email thread or a document, and sends that text to the AI provider along with your dictation ⁸⁶. Only switch it on for Modes that use a provider you trust with that content.

For a fully on-device setup, use a local transcription model and either no enhancement or VoiceInk Refine, which needs an Apple silicon Mac with 16 GB of memory ²³. Refine cannot follow custom prompts or use context ⁸⁷, so email drafting and rewrite Modes need Ollama or a cloud model.

History stays on the Mac until you delete it

VoiceInk's privacy policy says it does not store your transcriptions or audio on its servers ⁸⁸, and that it has no access to them unless you share them ⁸⁹. On the Mac, transcripts are kept indefinitely until you delete them ⁹⁰. Audio recordings are also kept indefinitely by default ⁹¹; automatic cleanup is optional, with a default of 7 days when switched on ⁹². If you dictate sensitive material, turn on cleanup. Once data reaches a cloud provider, that provider's own retention rules apply; the guide to voice data retention compares them.

How it compares with Superwhisper and Wispr Flow

VoiceInk is the cheapest and most inspectable of the three, Superwhisper covers the most devices with a similar local option, and Wispr Flow is the most polished cloud service. Which is better depends on where you work and whether your audio may leave your computer.

Against Superwhisper, the core technology is close: both run Whisper and Parakeet on the Mac, and Superwhisper made Parakeet available to its Pro users in July 2025 ¹⁰⁵. The differences are price, platforms and defaults. VoiceInk costs less and starts local; Superwhisper starts on its cloud models ⁹⁸ but runs on Windows and phones under one license. Read the Superwhisper guide and the question VoiceInk vs Superwhisper.

Against Wispr Flow, the difference is local versus cloud. Every Wispr Flow dictation is sent to its servers ⁹⁹, an internet connection is required ¹⁰², and its two data-sharing switches start on ¹⁰⁶. In return, you get no setup and the highest ratings of the three. If you want the same no-setup feel in VoiceInk, you need to configure an AI provider yourself. See the Wispr Flow guide and the question VoiceInk vs Wispr Flow. The comparison of AI dictation apps and MacWhisper cover the wider field.

Built-in Mac dictation may be enough

Before you pay for anything, try the dictation built into macOS. It is free, and Apple says general text dictation on a Mac is processed on the device rather than on Siri servers ¹⁰⁷. The Apple Speech model VoiceInk offers had a 31.34 percent error rate in loud cafe noise ⁵⁰ against 7.85 for Parakeet ⁴⁸, but on clean read English the gap was smaller, 6.73 ⁴⁴ against 4.48 ⁴². The guide to Apple Dictation privacy explains how to confirm the on-device setting, and the best dictation app for Mac compares the paid options.

Who VoiceInk suits

• Developers and privacy-minded Mac users who want to read the code and keep audio local.
• People who dislike subscriptions: one payment covers updates ⁹.
• Writers with accents or noisy rooms, because the default Parakeet model held up best in our tests.

It suits you less if you need Windows or Android, if you are on an Intel Mac, or if you want polished AI rewriting without setting up an API key. VoiceInk has no Windows version. Yap, which we make, is an on-device dictation app for Mac and Windows; our overview of dictation on Windows covers the options on a PC.

Setting up VoiceInk takes about five steps

These steps describe VoiceInk 2.x on macOS 15, as documented on October 27, 2026.

1. Download the app from tryvoiceink.com, or install it with brew install --cask voiceink ³¹, and move it to Applications.
2. Grant permissions when asked: Microphone to record, Accessibility to paste text, and optionally Screen Recording for screen context.
3. Follow onboarding, which downloads Parakeet V3 as the local starting model ¹³.
4. Choose Start 7-day Trial or activate a license ¹⁰.
5. Set a recording shortcut, place your cursor in any app, press the shortcut and speak.

To add a model, open AI Models, then Local, download it, and select it in a Mode ¹². To try AI cleanup without sending text anywhere, download VoiceInk Refine from the same tab if your Mac has 16 GB of memory ²³. If you later turn on a cloud provider, check what each Mode sends before you dictate anything confidential.