← Blog

Offline Dictation with Parakeet

Your voice, your device, your text. When the on-device model is loaded, Hovor's Parakeet model transcribes speech entirely on your Mac or iPhone and the audio is not uploaded. If the model is still loading — after an app relaunch, or while it is still downloading — that recording falls back to server transcription.

Last updated: July 19, 2026

What "on-device" actually means

"On-device" means Hovor's Parakeet speech model runs the entire transcription step directly on your Mac or iPhone's own chip — audio is captured, converted to text, and discarded locally, with no network call carrying it anywhere. One caveat matters: this holds once the model is loaded. If the model is still loading — after an app relaunch, or while it is still downloading — that recording falls back to server transcription. This differs from cloud dictation, where audio always uploads to Hovor's server and OpenAI's gpt-4o-mini-transcribe API before a transcript returns.

On-device Parakeet covers the speech-to-text step only. Turning that raw transcript into cleaned-up, styled text (grammar, punctuation, tone) is a separate stage that still needs Hovor's server, Apple Foundation Models on iOS 26+, or your own BYOK key — Hovor's fully local tone-formatting code exists but is not functional on any platform today.

Cloud dictationOn-device Parakeet
Where audio goesUploaded to Hovor's server, forwarded to OpenAI's gpt-4o-mini-transcribeOn-device when the model is loaded; server fallback while it is loading
Works without networkNoYes, after a one-time model download
Latency sourceNetwork round trip + server queue + API inferenceOn-device inference only (CPU on iOS; CPU + Neural Engine on macOS)
Cost modelCounts against the 2,000 words/week free tier, or Pro subscriptionOne-time Local Unlock purchase, no per-word metering
Device requirementAny device, no RAM or storage gatemacOS: none. iOS: ≈4 GB+ RAM, plus a server-side flag must be on

Why offline = private

Offline dictation is private because there is no server in the loop to breach, subpoena, or leak: audio processed on-device by Parakeet does not cross a network boundary, so there is nothing for a company, attacker, or legal request to intercept. This applies to recordings made once the model is loaded; a recording taken while the model is still loading falls back to server transcription. Cloud dictation instead sends audio to remote infrastructure, where confidentiality depends on encryption and written policy rather than on the audio simply never leaving your hands.

If you are dictating:

…the question "who can hear this?" has a simple answer: only you, once the model is loaded — the audio is not uploaded to Hovor's server or any other server. The one exception is the model-loading window itself, when that recording falls back to the server.

For users subject to professional confidentiality obligations (lawyers, doctors, therapists, journalists), on-device dictation is not just a preference. It is often the only architecture that satisfies their obligations.

Works without internet

Offline dictation keeps working with zero network connectivity because the Parakeet model, once downloaded, runs entirely on your device's CPU (iOS) or CPU and Neural Engine (macOS) — there is no round trip to fail. Cloud dictation, by contrast, stops the moment your connection drops, because its transcription step depends on reaching Hovor's server and OpenAI's API.

On a long-haul flight. On a train through a tunnel. In a basement office with poor signal. In a country with unpredictable mobile coverage. In a meeting room where IT policy blocks outbound traffic. Wherever cloud dictation fails or becomes unreliable, on-device Parakeet keeps transcribing at the same quality.

Once the model is loaded, there is no spinner and no "waiting for server" delay — it runs whether or not a router is nearby. The one window where a router still matters is right after an app relaunch or while the model is still downloading, when a recording falls back to server transcription.

Parakeet: the model Hovor runs locally

Hovor runs Parakeet TDT 0.6B v3, a speech-recognition model NVIDIA built and released under the CC-BY-4.0 licence (verified against NVIDIA's published model card, checked 2026-07-19); Hovor did not train it. The model supports 25 European languages, including Ukrainian, trained with balanced sampling across languages and a single 8,192-token tokenizer shared across all of them — no per-language switching required.

Ukrainian gets the same training weight as German or French, not an afterthought. The model handles apostrophes correctly (м'ясо, об'єкт), preserves soft signs (день, тінь), and follows Ukrainian capitalization rules rather than mistakenly applying English conventions.

On iOS, Parakeet runs CPU-only rather than using the Neural Engine — a deliberate trade-off tied to background-execution limits on iOS for processes like the keyboard extension. On macOS it runs with both CPU and Neural Engine, which keeps it fast and battery-efficient.

Hovor selects and ships Parakeet; it does not fine-tune or modify it. The quality reflects NVIDIA's engineering on the model, and Hovor's engineering on the pipeline around it.

Turning on offline dictation (and Local Unlock)

On-device Parakeet ships on macOS unconditionally, and on iPhones with roughly 4 GB of RAM or more — a server-side flag can also disable it independently of your hardware. It requires Local Unlock, a one-time $49.99 / 1,999 UAH purchase, family-shareable, sold through the App Store on iOS and macOS, and separately via Monobank on the web — no subscription.

PlatformOn-device ParakeetRequirement / status
macOSYesAvailable on every Mac Hovor supports, no RAM gate
iOSYes, conditional≈4 GB+ RAM device, and the server-side parakeet_enabled flag must be on
WindowsNoTauri rewrite in progress; no on-device dictation shipped yet

Free-tier cloud dictation (2,000 words per week) uses Hovor's server and works without Local Unlock. Local Unlock also includes unlimited dictionary replacements, snippets, and custom workflows; it is a single purchase that covers all the features that run entirely on your device.

To switch to on-device mode after purchasing:

  1. Open Hovor on your Mac or iPhone.
  2. Go to Settings → Dictation → Speech model.
  3. Select Parakeet (on-device).
  4. Hovor will download the model if it is not already on your device. This is a one-time download over Wi-Fi; after that, no connection is needed.

Good to know: model download and device requirements

Parakeet downloads once over Wi-Fi and then needs no connection for transcription, including model updates you trigger manually from Settings; the download requires a modest amount of local storage, shown before you start. On-device Parakeet covers speech-to-text only — text formatting still needs the server, Apple Foundation Models (iOS 26+), or your own BYOK key, since Hovor's fully local tone-formatting path is not shipped.

Frequently asked questions

Does offline Parakeet dictation also format and clean up my text without internet?

No, not fully. On-device Parakeet handles only the speech-to-text step: once the model is loaded, your voice becomes a raw transcript entirely on your device, with no network call. If the model is still loading — after an app relaunch, or while it is still downloading — that recording falls back to server transcription. Turning that raw transcript into formatted, toned final text (grammar, punctuation style, casual vs. formal) is a separate stage that still needs either Hovor's server, Apple Foundation Models on iOS 26+ (also on-device, but a distinct feature from Parakeet), or your own BYOK key. Hovor also ships code for a fully local LLM-based tone formatter, but it is not functional on any platform today — real tone formatting always runs through one of the three paths above. If you dictate with no network at all and don't have Apple Foundation Models enabled, you get the raw Parakeet transcript without automatic formatting.

What happens to my audio when I use on-device Parakeet transcription?

When the on-device model is loaded, transcription runs entirely on your device and the audio is not uploaded. If the model is still loading — after an app relaunch, or while it is still downloading — that recording falls back to server transcription. With the model loaded there is no network call carrying your voice, so there is no server log to audit and no third-party infrastructure in the path. Hovor's cloud dictation, by contrast, always sends audio to Hovor's server and onward to OpenAI's gpt-4o-mini-transcribe API to produce a transcript.

Which Hovor platforms support on-device Parakeet dictation?

Only macOS and iOS. On macOS, Parakeet is available unconditionally on every Mac Hovor supports. On iOS, it requires roughly 4 GB of device RAM or more, and it can also be switched off server-side independently of your hardware via a feature flag. Hovor's Windows client is a Tauri-based rewrite still in progress; it does not yet ship any on-device dictation.

How much RAM does my iPhone need for offline dictation?

Roughly 4 GB of physical memory or more. Hovor checks your iPhone's memory against a 4,000,000,000-byte threshold before showing the on-device Parakeet option in Settings; iPhones with less RAM won't see it at all. This opens the option up to iPhone 11, SE (2nd/3rd gen), 12 series, 13 series, and every 6 GB/8 GB device, while excluding only the oldest, lowest-memory models. On top of the RAM gate, a server-side flag can also disable Parakeet independently of your device's hardware. macOS has no RAM gate — it is available on every Mac Hovor supports, regardless of memory.

Does Local Unlock require a subscription?

No. Local Unlock is a one-time, non-consumable purchase — $49.99 or 1,999 UAH — sold through the App Store on iOS and macOS, and separately via Monobank on the web. It's family-shareable and never expires, and it covers on-device Parakeet dictation plus unlimited dictionary replacements, snippets, and custom workflows, all in a single purchase. Hovor's free tier (2,000 cloud words per week) and Pro subscription remain separate, unaffected by whether you've bought Local Unlock.

Which languages does Parakeet support, and is Ukrainian handled well?

Parakeet TDT 0.6B v3 supports 25 European languages — English, Ukrainian, Spanish, French, German, Italian, Portuguese, Dutch, Polish, Russian, Czech, Slovak, Slovenian, Croatian, Bulgarian, Romanian, Hungarian, Greek, Finnish, Swedish, Danish, Estonian, Latvian, Lithuanian, and Maltese — using a single tokenizer and balanced training sampling across all of them, per NVIDIA's published model card. Ukrainian gets the same training weight as German or French, not an afterthought treatment. It correctly handles Ukrainian apostrophes (м'ясо, об'єкт), soft signs (день, тінь), and native capitalization rules rather than defaulting to English conventions. Languages outside this list — such as Japanese, Chinese, Korean, Arabic, or Hindi — are handled by Hovor's separate Whisper on-device model instead, not by Parakeet.

Dictate privately, without the internet

Download Hovor for Mac or iPhone. Free tier includes 2,000 cloud words/week. Local Unlock adds unlimited on-device Parakeet: one-time purchase, family-shareable, no subscription.

Get Hovor