What is Hold Speak?
Hold Speak is a free, local push-to-talk dictation app for macOS. You hold a hotkey, say what you want to type, and release — the recognized text is typed straight into whatever text field has focus, in any app: Slack, Cursor, VS Code, a terminal, ChatGPT or Claude in the browser, Notes, Telegram. Speech recognition runs on your Mac with NVIDIA's Parakeet Ultra through FluidAudio on the Neural Engine (the default on macOS 14+ with Apple Silicon) or OpenAI's Whisper models through WhisperKit on the GPU and Neural Engine — so by default there is no cloud service, no account, no API key and no subscription. If you want, you can switch to OpenAI or Google Gemini with your own API key instead.
It was built for one job: talking to AI agents faster than you can type to them. Typing averages 40–60 words per minute; comfortable speech is 130–160. When most of your work is writing prompts, that gap is the bottleneck. Hold Speak lives in the menu bar and stays out of the way until you hold the key.
How does it work?
- Hold the push-to-talk key — Right Option by default, or any other key you pick. A small HUD shows the live mic level.
- Speak. Audio is kept in memory only; it is never written to disk.
- Release. Whisper transcribes the clip on-device (or OpenAI or Gemini in the cloud, if you chose one). Fillers like “ummm” are trimmed and your terminology dictionary fixes names and jargon.
- The text appears in the focused field, typed as keyboard input — your clipboard is left alone and password fields are skipped.
Short taps under 150 ms don't trigger recording, so the key still works as a normal modifier. Recent transcriptions stay in the menu bar popover in case the cursor was in the wrong place.
Why local dictation?
- Private by default. With the local engines (Parakeet Ultra, the default on macOS 14+, or Whisper Turbo) your voice never leaves the Mac. The optional OpenAI and Gemini engines send audio to OpenAI or Google only if you switch to one. No telemetry, no analytics, no crash-reporting SDK. Details on the privacy page.
- Works offline. With Whisper or Parakeet, after a one-time model download, dictation works on a plane, on bad Wi-Fi, or with all outbound traffic blocked.
- Cloud accuracy when you want it. Bring your own OpenAI or Google Gemini API key for noticeably better results on jargon and mixed languages — OpenAI GPT Transcribe costs about $0.27 per hour of speech, billed to your key. Turn on Fallback and a dictation the cloud couldn't handle (no credits, no internet) goes to the other engine or to the local model you picked last on your Mac (OpenAI Whisper Turbo or NVIDIA Parakeet Ultra).
- Free, with no limits. No per-minute billing, no word caps, no trial that runs out.
- Fast. With NVIDIA Parakeet Ultra (the default on macOS 14+, Apple Silicon) text is ready about 0.1–0.2 s after you release the key; OpenAI Whisper Turbo (large-v3 distilled), the default on macOS 13, takes about 2.5–3.5 s. In our test on an M2 Max with synthetic speech, Parakeet was as accurate or slightly better, including Russian with English IT terms.
- Built for mixed-language tech speech. Russian or Ukrainian with English terms stays readable — “pull request” doesn't turn into a transliteration. A built-in dictionary with 110–130 IT terms per language (English, Russian, Ukrainian) is fully editable.
How does it compare to other Mac dictation apps?
Hold Speak does one thing — hold, speak, release, text appears — and does it free and on-device by default. Paid apps add AI rewriting, meeting recording or file transcription on top. Prices as of September 2026, from each vendor's site:
| Hold Speak | Superwhisper | Wispr Flow | Apple Dictation | |
|---|---|---|---|---|
| Price | Free | Free tier; Pro $8.49/mo or $249.99 lifetime | 2,000 words/week free; Pro from $12/mo | Free, built in |
| Where speech is processed | On your Mac by default; optional OpenAI or Gemini cloud with your own key | Local or cloud | Cloud only | On-device or Apple servers, depends on settings |
| Works offline | Yes, with the local engines (Parakeet or Whisper) | Yes, with local models | No | Partly |
| Hold-to-talk into any app | Yes | Yes | Yes | Toggle; stops after 30 s of silence |
| Custom vocabulary | Yes, per language | Yes | Yes | Not documented |
| Source code public | Yes, MIT license | No | No | No |
More apps — MacWhisper, VoiceInk, Aiko, Spokenly — with sources: Hold Speak as a free Superwhisper alternative.
Requirements
- A Mac with Apple Silicon (M1 or later) and macOS 13 Ventura or later.
- Disk space for one local model: OpenAI Whisper Turbo 1.5 GB, NVIDIA Parakeet Ultra 610 MB. Parakeet needs macOS 14 or later on Apple Silicon. A Whisper model already downloaded by MacWhisper or another WhisperKit app is reused.
- Three macOS permissions: Microphone, Accessibility (to type the text) and Input Monitoring (for the global hotkey).
- Languages: auto-detect, or one of 16 fixed languages — English, Spanish, German, French, Italian, Portuguese, Dutch, Polish, Russian, Ukrainian, Turkish, Arabic, Hindi, Chinese, Japanese, Korean.
FAQ
Is Hold Speak really free?
Yes. It's free to download and use, with no subscription, no paid tier and no usage limits. The only thing that can cost money is an optional cloud engine, billed directly to your own API key: about $0.27 per hour of speech with OpenAI GPT Transcribe, $0.30 with Gemini 3.5 Transcribe. It's open source under the MIT license, with the code on GitHub.
Does it work without internet?
Yes, with the local engines — Parakeet Ultra (the default on macOS 14+) or Whisper Turbo. The only download is the model on first use. After that, dictation needs no network at all. The optional OpenAI and Gemini engines need internet.
Does my voice get sent anywhere?
Not by default: Parakeet or Whisper transcribes on your Mac and audio is never uploaded or saved to disk. Only if you switch the engine to OpenAI or Gemini is each dictation sent to that provider with your own API key. Details on the privacy page.
Which apps can I dictate into?
Any macOS app with a text field. Hold Speak types the text as keyboard input, so it works in browsers, chat apps, editors, IDEs and terminals.
Why does macOS say the app is blocked?
The app is self-signed rather than notarized through the App Store. Open System Settings → Privacy & Security and click Open Anyway once — the install guide walks through it.
How is it different from Apple's built-in Dictation?
Hold Speak uses Whisper or Parakeet models, keeps English technical terms intact in mixed-language speech, lets you teach it your own vocabulary, and is driven by a single hold-to-talk key.
How do I make it faster?
On macOS 14+ with Apple Silicon, switch the model to NVIDIA Parakeet Ultra in Preferences → Audio → Model: text is ready in about 0.1–0.2 s instead of 2.5–3.5 s with OpenAI Whisper Turbo. It covers 25 European languages, including English, Russian and Ukrainian; for other languages, or on macOS 13, stay with Whisper Turbo.