Skip to content
Petal

How-toMac6 min read

Whisper App for Mac: Run Whisper Locally for Dictation

Use a Whisper app for Mac to run Whisper locally on Apple silicon. Set up Whisper Large V3 Turbo in Petal and compare Parakeet and Qwen3-ASR.

Published by Optimal Apps, the maker of Petal.

On this page

A Whisper app for Mac runs OpenAI's open Whisper speech model on your own computer, so your audio does not go to a server. On Apple silicon, Whisper runs well through Core ML with argmax's WhisperKit, through whisper.cpp, or through MLX. If you want Whisper for dictation in any app, Petal includes Whisper Large V3 Turbo through WhisperKit: download it in Settings > Models, select it, press Option-Space, and speak. This guide explains local Whisper, shows the Petal setup, and compares Whisper with the other models in Petal.

The steps match Petal 2.8.1 on macOS 26. Optimal Apps publishes Petal, which is free and open source.

Quick checklist

Check Why
The Mac has Apple silicon Local Whisper needs the Neural Engine and GPU for usable speed
About 1.1 GB of free disk space The size of Whisper Large V3 Turbo in Petal
The model finished its download Petal cannot use a partial model
The spoken language is set, or left on detect Whisper can detect it, but a set language avoids mistakes
Microphone and Accessibility are on Needed to record and to paste

How local Whisper works on Apple silicon

Whisper is a family of open speech models from OpenAI. The original code is in Python and runs best on NVIDIA GPUs. On a Mac, three projects make it fast:

  • WhisperKit by argmax converts Whisper to Core ML, so it runs on the Apple Neural Engine. Petal uses WhisperKit.
  • whisper.cpp is a C/C++ port that uses Metal on the GPU. It is a command-line tool, so it suits scripts more than dictation.
  • MLX is Apple's machine learning framework. Several Whisper ports run on it.

Whisper Large V3 Turbo is a smaller, faster version of Whisper Large V3. It has 809M parameters, against 1.55B for the full Large V3, with close to the same accuracy in many languages.

Local Whisper speed depends on the chip, the memory, the length of the audio, and the model. We did not publish timing numbers, because one Mac does not predict another. Test with your own voice on your own Mac. Read the same paragraph into two models, and compare the wait after you stop and the number of words that you must fix.

Run Whisper from the command line

If you only want to transcribe a file once, whisper.cpp works:

  1. Install it with Homebrew: brew install whisper-cpp.
  2. Download a model file from the whisper.cpp project.
  3. Run the command-line tool on a WAV file.

This is fine for files. For daily dictation, you need an app that records with a shortcut and pastes into the front app.

Set up a Whisper app for Mac: Petal step by step

Step 1: Install Petal

  1. Download Petal from GitHub Releases, or from the Petal for Mac page.
  2. Drag Petal to Applications and open it.
  3. Complete the setup: allow the microphone and Accessibility, and keep the default Option-Space shortcut or set your own.

Step 2: Download Whisper Large V3 Turbo

  1. Click the Petal icon in the menu bar and choose Petal Settings….
  2. Open Models.
  3. Find Whisper Large V3 Turbo. Its card says "High-quality transcription in 99 languages" and shows the size, about 1.1 GB.
  4. Click Get and wait for the download.
  5. Click the card to select it.

The first dictation after a model change can take longer, while the model loads. The card says Warming up during this time.

Step 3: Set the spoken language

Whisper Large V3 Turbo and Qwen3 ASR 1.7B can take a set language, instead of detecting it.

  1. Open Settings > Recording and find Spoken Language.
  2. Choose your language, or leave it on Automatic.

A set language helps with short dictations, where there is little audio to detect the language from.

Step 4: Dictate

  1. Click in any text field.
  2. Press Option-Space, speak, and press it again. Or hold it, speak, and release.
  3. Petal transcribes with Whisper on the Mac and pastes the text.

To also remove filler words, select Petal W1 in Settings > Intelligence. It runs on the Mac too.

Whisper vs the other models in Petal

Petal rates each model with a speed meter and a quality meter from 1 to 5. These are Petal's own ratings, not a benchmark.

Model Engine Size Languages Speed Quality
Parakeet 110M FluidAudio (Core ML) ~455 MB English 5 2
Parakeet 0.6B V3 FluidAudio (Core ML) ~485 MB 25 European 5 4
Parakeet Unified 0.6B FluidAudio (Core ML) ~625 MB English, live transcript 5 3
Nemotron Unified 0.6B FluidAudio (Core ML) ~1.2 GB 25 European, live transcript 5 4
Whisper Large V3 Turbo WhisperKit (Core ML) ~1.1 GB 99 2 5
Qwen3 ASR 1.7B MLX ~2.5 GB 30, auto detect 3 5
Voxtral Realtime 4B MLX ~3.2 GB 13 4 5
Apple Speech Apple Speech framework No download System languages 5 3

How to choose:

  • English dictation, all day: Parakeet 110M. It is the default for this reason.
  • European languages: Parakeet 0.6B V3, or Nemotron Unified for a live transcript while you speak.
  • A language that Parakeet does not cover: Whisper Large V3 Turbo.
  • The most accurate result: Qwen3 ASR 1.7B. Petal's description calls it the most accurate open model on the Open ASR Leaderboard.

Use Whisper for files and Parakeet for dictation

You do not have to pick one. Petal can use one model for dictation and another for audio files.

  1. Download both models.
  2. Open Settings > Router and find Speech Models.
  3. Set Dictation to Parakeet 110M.
  4. Set Audio Files to Whisper Large V3 Turbo.

Then drop a recording on the menu bar icon to transcribe it with Whisper. Read how to transcribe an audio file on Mac.

Common problems with local Whisper

  • The first dictation is slow. Core ML compiles and loads the model the first time. The card says Warming up. Later dictations start faster.
  • Whisper adds words that you did not say. Whisper can add text to silence, such as "Thank you." Turn on Trim Silence in Settings > Recording, which removes quiet parts before transcription.
  • The wrong language appears. On a short dictation, Whisper can detect the wrong language. Set Spoken Language in Settings > Recording.
  • The paste does not arrive. This is a permission problem, not a model problem. Read accessibility permission not working on Mac.

Is local Whisper private

Yes, when the model runs on your Mac. Petal records the audio, runs Whisper on the Mac, and pastes the text. The audio does not go to OpenAI or to Optimal Apps. If you turn on a Cloud Model for cleanup, only the text goes to the provider that you choose. For that setup, read per-app dictation cleanup.

Other Whisper apps for Mac

  • MacWhisper focuses on file transcription with timestamps and exports. Dictation is only in the website edition. Read MacWhisper alternatives.
  • Superwhisper has free local Whisper models and paid cloud models, with custom modes.
  • VoiceInk is open source under GPL v3, with a paid signed build.

For a wider view, read the best dictation apps for Mac and offline speech to text for Mac.

Questions

Can I run Whisper locally on a Mac?

Yes. On Apple silicon, Whisper runs on the Mac through Core ML with WhisperKit, through whisper.cpp, or through MLX. Apps such as Petal, MacWhisper, and Superwhisper package this for you.

Is Whisper Large V3 Turbo good for dictation?

It is accurate and supports 99 languages, but it is slower than Parakeet models. For short English dictation, Parakeet is faster. Whisper is a good choice for other languages and for audio files.

Is there a free Whisper app for Mac?

Yes. Petal is free and open source and includes Whisper Large V3 Turbo as a model choice. MacWhisper and Superwhisper also have free plans with Whisper models.

Does Whisper on Mac work offline?

Yes. After you download the model, Whisper runs without the internet.