Memos, attendance notes, and client correspondence take longer to type than to say. Hovor turns speech into formatted text on iOS and macOS, with a choice between server-side processing and on-device processing, and a custom dictionary for case names and jurisdiction-specific terms. This page describes exactly where your data goes at each stage.
Hovor records speech, converts it to text, and applies a formatting pass — punctuation, capitalization, paragraph breaks, optional tone — before pasting the result into whatever application has focus: a document, an email client, a case management field, a chat window. Speech-to-text and formatting are separate stages, and each can run on-device or remotely.
On iOS this happens through the Hovor keyboard extension; on macOS it happens through system-level text injection into the active app. Which combination you pick for the two stages determines the full data path for a given dictation — the table below maps every stage to its actual destination.
Speech-to-text and formatting are separate stages with separate routing. Cloud mode sends both to Hovor's server, which forwards audio and text to OpenAI. On-device transcription (Parakeet) keeps audio local, with a load-time server fallback covered below. On-device formatting works only via Apple Foundation Models, iOS-only. BYOK reroutes formatting to a provider you choose, using your own key, bypassing Hovor's server.
| Pipeline stage | On-device | Hovor server | Third-party (BYOK) |
|---|---|---|---|
| Speech-to-text (cloud/default) | No | Yes — audio forwarded to OpenAI | No |
| Speech-to-text (Parakeet, on-device) | On-device once the model is loaded | Fallback while the model is loading | No |
| Formatting/cleanup (default) | No | Yes — text forwarded to OpenAI (gpt-4o-mini) | No |
| Formatting (Apple Foundation) | Yes — iOS 26+, iPhone 15 Pro or newer, 8 GB+ RAM, Apple Intelligence enabled | No | No |
| Formatting (BYOK) | No | No — bypassed | Yes — direct to your chosen provider with your key |
| Custom dictionary (case names, terms) | No | Yes — stored on server, synced to devices | No |
To keep an entire dictation session (audio, transcript, formatted text) off Hovor's server, you need Parakeet for speech-to-text plus either a qualifying Apple Foundation Models device or BYOK for formatting. Using the default configuration on either platform sends both audio and text through Hovor's server to OpenAI. Dictionary entries themselves are stored on Hovor's server regardless of which STT or formatting mode is active, because dictionary sync is a separate system from the transcription pipeline.
Hovor's on-device speech-to-text runs on the Parakeet model (NVIDIA, CC-BY-4.0, trained across 25 European languages) on iOS and macOS. On iOS, the on-device option is gated to devices with roughly 4 GB of physical RAM or more; on macOS it is available unconditionally, regardless of device.
iPhone 11 and every newer mid-tier or Pro model at 6 GB+ RAM qualifies on iOS. On-device transcription already works on the free tier under the same weekly quota as cloud; Local Unlock — a one-time $49.99 / 1,999 UAH purchase, family-shareable, that also removes limits on dictionary entries, snippets, and custom workflows — or a Pro subscription removes that cap. On a device below the memory threshold, speech-to-text runs through Hovor's server by default, which forwards the audio to OpenAI's gpt-4o-mini-transcribe.
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.
A common assumption is that "on-device dictation" means the whole pipeline stays local. It does not, on any current Hovor build. Formatting has exactly three working backends: Hovor's server (OpenAI gpt-4o-mini), Apple Foundation Models, and BYOK (your own OpenAI, Anthropic, or custom-endpoint key). A fourth path, a local EuroLLM model, exists in the codebase but does not work on any platform.
Apple Foundation Models formatting requires iOS 26 or later, an iPhone 15 Pro or newer, at least 8 GB of physical RAM, and Apple Intelligence enabled in Settings — it is iOS-only and is not available on macOS at all. Falling short of any one of those checks routes formatting to Hovor's server instead, with the same OpenAI-based path as the default.
Practically: a lawyer on macOS, or on an iOS device that fails any of those checks, using Parakeet for transcription but not BYOK, has on-device audio capture but the formatted output still travels to Hovor's server. Only a qualifying iOS device with Apple Foundation Models, or any platform with BYOK configured, keeps the formatting stage off Hovor's infrastructure as well.
Dictation mishears proper nouns it has not seen often in training data: party names, opposing counsel, case citations, jurisdiction-specific Latin terms, statute names. Hovor's custom dictionary maps variant transcriptions (the phonetic mishears) to a canonical spelling, and feeds those same entries to the formatting model as vocabulary context, so the model applies the correct spelling consistently rather than only via string substitution after the fact.
Setup is the same mechanism used across Hovor's dictionary feature generally: list the mangled variants separated by |, then the correct output, for example Sotheby's|so the bees → Sotheby's or a client or opposing-party name that speech models otherwise flatten to a common-word substitute. Entries sync across your devices through Hovor's server within seconds, so an entry added on your phone during a commute is present on your desktop by the time you sit down to draft. Free tier includes the dictionary with a limited entry count; Local Unlock removes the entry limit.
A single recording sent to Hovor's server, whether via the streaming or standard endpoint, is capped at 31 minutes (1,860 seconds) per request; past that, the server rejects the request with a recording_too_long error rather than truncating it silently. For a memo or attendance note longer than that in one sitting, split it into separate recordings, each of which is transcribed and formatted independently.
On macOS, continuous mode avoids the question entirely for most drafting sessions: trigger it once, and each pause in your speech ends a phrase that is transcribed, formatted, and inserted into the active document immediately, then listening resumes for the next phrase. A multi-paragraph memo accumulates as a sequence of clean insertions rather than requiring one uninterrupted recording. On iOS, dictation goes through the Hovor keyboard extension: you record, release, and the formatted chunk is inserted; there is no continuous hands-free mode on iOS in the current release.
The same dictated content often needs a different register depending on destination: a client letter reads differently from an internal file note. Hovor's formatting pass supports selectable tone styles that adjust register — a more formal style for correspondence versus a plainer style for internal notes — without changing the substance of what you said.
Tone selection happens before you dictate and applies during the formatting stage itself, so it follows the same on-device/server/BYOK routing as any other formatting: a tone applied via Hovor's server sends the text to OpenAI along with the tone instruction, while a tone applied via Apple Foundation Models or BYOK keeps that same text off Hovor's server.
Drafting memos, correspondence, and notes by keyboard for hours at a stretch is a documented aggravator for repetitive strain injury and general typing fatigue. Dictation replaces the composition phase — the longest and most repetitive part of drafting — with speech, while editing and review remain keyboard-based but shorter.
This is a mechanical description of typing load, not a medical claim: whether dictation meaningfully reduces RSI risk for a given individual depends on that individual's overall typing volume, posture, and existing condition.
Free tier includes 2,000 words per week of cloud dictation, cloud transcription, and the dictionary with a limited entry count. Pro ($11.99/month or $89.99/year, up to 5 devices) removes the word limit but does not change where data is processed — cloud dictation still routes through Hovor's server.
Local Unlock ($49.99 / 1,999 UAH, one-time, family-shareable) enables on-device Parakeet transcription and removes dictionary/snippet/workflow entry limits. BYOK Unlock ($24.99 / 999 UAH, one-time) enables routing formatting, and translation where supported, to your own OpenAI or Anthropic key instead of Hovor's server.
The Local + BYOK Bundle ($69.99 / 2,799 UAH) combines both at a discount to buying separately. None of these purchases change what claims Hovor makes about compliance; they change which infrastructure processes your data.
Both modes exist, and you choose. In cloud mode, audio is sent to Hovor's server, which forwards it to OpenAI for transcription and to OpenAI again for punctuation and formatting. In on-device mode (Parakeet, iOS with 4 GB+ RAM or macOS), transcription runs entirely on your device once the on-device model is loaded 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. Formatting can also run on-device, but only via Apple Foundation Models, which require iOS 26 or later, an iPhone 15 Pro or newer, at least 8 GB of RAM, and Apple Intelligence enabled; short of that, formatting goes to a server instead. This is a description of data flow, not a compliance determination — each firm should assess whether either mode meets its own confidentiality obligations.
Only if you use on-device transcription and on-device formatting together, and only for recordings where the on-device model was already loaded — a recording made while the model is still loading falls back to server transcription regardless of your settings. Hovor's custom dictionary (case names, party names, jurisdiction-specific terms) is stored on Hovor's server and synced to your devices, regardless of which STT mode you use, so dictionary text itself reaches the server even when audio does not. If you want the full session (audio, transcript, and formatting) to stay off Hovor's server, use Parakeet for speech-to-text and Apple Foundation Models for formatting on an iPhone meeting all four requirements: iOS 26 or later, iPhone 15 Pro or newer, at least 8 GB of RAM, and Apple Intelligence enabled. Apple Foundation Models formatting is not available on Mac.
BYOK (Bring Your Own Key) lets you supply your own OpenAI or Anthropic API key, stored in the device Keychain, so formatting requests go directly from your device to that provider using your key, never proxied through Hovor's server. It changes who your text is sent to (your chosen provider instead of Hovor's server) but it does not eliminate a third party from the pipeline. BYOK Unlock is a one-time $24.99 / 999 UAH purchase. Whether this arrangement satisfies a given confidentiality or ethical-duty obligation is a determination each firm has to make itself.
Hovor is a general-purpose voice-to-text app for iOS and macOS with a custom dictionary, tone-based formatting, and an on-device transcription option; it has no legal-specific workflow templates, no case-management integration, and no certification aimed at the legal industry. Whether it fits a given practice's workflow and confidentiality requirements as well as, better, or worse than a dedicated legal dictation product depends entirely on that practice's own requirements. We have not independently verified current Dragon Legal, BigHand, or Winscribe pricing or features and make no comparison claim here.
A single recording sent to Hovor's server is capped at 31 minutes (1,860 seconds); a request past that limit is rejected with a recording_too_long error rather than truncated. On macOS, continuous mode removes the need to fit a long memo into one take: each pause ends a phrase, which is transcribed, formatted, and inserted immediately, so a multi-paragraph memo accumulates as a sequence of clean insertions rather than one long recording.
Hovor for iOS and macOS: cloud or on-device transcription, custom dictionary for case names and terminology, and BYOK for formatting with your own key. Free tier includes 2,000 words/week.
Get Hovor