A Hovor workflow is a saved combination of speech model, tone style, and translation step. Switch workflows and you switch all three at once — plain transcription for quick notes, a casual style for chat, a formal style for email — without touching a settings screen mid-sentence.
A workflow bundles four independent settings into one switchable unit: the speech-to-text model that transcribes your voice, an optional tone style that reformats the raw transcript, an optional translation step with a target language, and behavior flags (whether your custom dictionary applies, whether streaming/realtime mode is used). Each of the three model stages can point at a different backend — Hovor's server, an on-device model, or your own API key — independently of the others.
Because these are saved as named, switchable pipelines rather than one global setting, you can keep a "Default" workflow for fast, literal notes and a "Formal" workflow for client correspondence, and jump between them per recording instead of reconfiguring Settings every time the context changes.
Hovor seeds every install with three workflows: Default (speech-to-text only — no tone step runs, so output is the most literal and fastest to produce), Casual (applies the casual tone style, suited to everyday messages), and Formal (applies the formal tone style, suited to professional writing). All three can be renamed, re-pointed to a different speech model, or deleted; deleting a built-in workflow tombstones it so it doesn't silently reappear on the next launch.
Earlier catalog versions also shipped Email, Slack, Bullet, Coding, Biblical, and McKinsey seed tone styles as starting points for custom workflows. A 2026 catalog-slimming migration removed the ones most users never edited or referenced, to keep the default catalog small — any workflow a user had actually built on top of those styles was preserved untouched.
The table below lists the complete set of backends for each of the three workflow stages — speech-to-text, tone, and translate — exactly as wired in Hovor's routing code, not as a roadmap. A blank platform requirement means the backend needs nothing beyond having Hovor installed.
| Stage | Backend | Platform / requirement | Works today |
|---|---|---|---|
| Speech-to-text | Hovor server (gpt-4o-mini-transcribe) | Any platform, internet required | Yes |
| Speech-to-text | Parakeet (on-device) | iOS/macOS, ≥4 GB RAM, free tier (capped) or Local Unlock/Pro (unlimited) | Yes |
| Speech-to-text | Whisper (on-device) | iOS/macOS, ≥6 GB RAM, free tier (capped) or Local Unlock/Pro (unlimited) | Yes |
| Speech-to-text | OpenAI / Deepgram / ElevenLabs (BYOK) | Any platform, your API key, BYOK Unlock or Pro | Yes |
| Tone / formatting | Hovor server | Any platform, internet required | Yes |
| Tone / formatting | Apple Foundation Models | iOS 26+ / macOS 26+, English only, ~8 GB RAM | Yes |
| Tone / formatting | OpenAI / Anthropic / Google / custom endpoint (BYOK) | Any platform, your API key or none for custom endpoints, BYOK Unlock or Pro | Yes |
| Tone / formatting | Local EuroLLM (on-device LLM) | Code present, not shipped as working | No |
| Translation | Hovor server | Any platform, internet required | Yes |
| Translation | Apple Translation | iOS 26+ / macOS 26+, no RAM floor, 19 languages | Yes |
| Translation | OpenAI / custom endpoint (BYOK) | Any platform, your API key, BYOK Unlock or Pro | Yes |
| Translation | Anthropic / Google selected for translate | Falls back to Hovor server for this stage | No (routes to server) |
| Translation | Local EuroLLM (on-device) | Code present, not shipped as working | No |
The pattern across all three stages: server is the universal fallback, on-device options exist only where a real model ships and a RAM floor is met, and BYOK options exist wherever a key can be supplied. Nothing silently routes to a non-functional local LLM — the code paths that would have done that were removed after evaluation.
Apple Foundation Models format tone on-device on iOS 26+ and macOS 26+, but only for English text and only on devices with roughly 8 GB of RAM or more — both are hard requirements in Hovor's provider capability table, not soft recommendations. Apple Translation has no RAM floor and covers 19 languages, but still needs iOS 26+ / macOS 26+.
Those 19 Apple Translation languages are English, Ukrainian, Russian, Polish, German, Spanish, French, Italian, Portuguese, Dutch, Arabic, Chinese, Japanese, Korean, Turkish, Indonesian, Thai, Vietnamese, and Hindi. If a device or OS version doesn't meet a gate, Hovor doesn't error out mid-dictation — the workflow's format or translate call falls back to the server backend for that run, and picks up the on-device path again automatically once the device qualifies.
Beyond the three built-ins, you can create a custom workflow: pick a source language, a speech-to-text model, an optional tone style (built-in or your own custom prompt), an optional translation target, and behavior flags, then give it a name and icon. Custom workflows are stored per-user and sync the same way built-ins do; nothing about a custom workflow is second-class compared to a shipped one.
A common pattern is one workflow per context: "Slack" pinned to a casual tone and no translation, "Client email" pinned to formal tone plus your own OpenAI key, "Ukrainian notes" pinned to Apple Translation targeting English. Each is a single tap to switch to before you start recording.
Every backend a workflow touches resolves to one of three entitlement channels: server (free tier's 2,000 words/week, or unlimited with Pro Monthly $11.99 / Pro Yearly $89.99), on-device (works free-tier under the weekly quota, unlimited with Local Unlock $49.99 / 1,999 UAH one-time or Pro), and BYOK (Pro, or the standalone BYOK Unlock, $24.99 / 999 UAH one-time). Before a workflow runs, Hovor checks every stage's channel against your entitlements and blocks the run at the first uncovered stage.
Local Unlock also covers unlimited dictionary replacements, snippets, and custom workflows in one purchase. A blocked run routes you to the matching paywall rather than executing partially or silently substituting a different backend — see how the same Local Unlock purchase enables fully offline speech-to-text in Offline Dictation with Parakeet.
A workflow is a named pipeline that bundles four independent choices: which speech-to-text model transcribes your voice, which tone style (if any) reformats the raw transcript, whether a translation step runs before or after formatting, and behavior flags like whether your custom dictionary is applied. Switching workflows switches all of that at once, so you can keep one workflow tuned for quick notes and a different one tuned for client email, without re-configuring settings each time.
Hovor ships three built-in workflows: Default (transcription only, no tone step — the fastest, most literal option), Casual (applies the casual tone style for everyday messages), and Formal (applies the formal tone style for professional writing). Earlier versions also shipped Email, Slack, Bullet, Coding, Biblical, and McKinsey seed styles; a 2026 catalog-slimming pass removed the ones most users never touched, and any workflow built on top of them stays intact so no one's setup silently breaks. All three built-ins, and any custom workflow you add, can still be edited or deleted per user.
Yes. A workflow's tone step can point at OpenAI, Anthropic, Google, or a custom OpenAI-compatible endpoint (self-hosted Ollama, LM Studio, vLLM) instead of Hovor's server, once BYOK is active — via a Pro subscription or the standalone BYOK Unlock ($24.99 / 999 UAH one-time) — and a key is saved in Keychain. If a BYOK call fails, Hovor can automatically fall back to the server tone backend rather than losing the run — that behavior is a toggle, not forced.
No local LLM tone-formatting backend is functional on any Hovor platform today. The only working tone backends are Hovor's server, Apple Foundation Models (iOS 26+ and macOS 26+, English only, requires roughly 8 GB of device RAM), and BYOK providers. A fully on-device EuroLLM tone backend exists in the codebase but was abandoned on iOS (no CPU-only model reached usable formatting quality) and was never validated on macOS — it should not be treated as an available feature.
Partially. Apple Translation runs fully on-device on iOS 26+ / macOS 26+ with no RAM floor and covers 19 languages including English, Ukrainian, Polish, German, French, and Spanish. OpenAI BYOK and a custom OpenAI-compatible endpoint also handle translation, routed through your own key. Anthropic and Google are valid tone-formatting BYOK providers but do not run the translate step themselves — a workflow with Anthropic or Google selected for translation falls back to Hovor's server for that step.
Each stage of a workflow (speech-to-text, tone, translate) resolves to one of three backend channels — Hovor's server, on-device, or BYOK — and each channel has its own entitlement: server access follows the free 2,000 words/week tier or a Pro subscription, on-device already works on the free tier under that same weekly quota and becomes unlimited with Local Unlock ($49.99 / 1,999 UAH one-time) or Pro, and BYOK works with either a Pro subscription or the standalone BYOK Unlock ($24.99 / 999 UAH one-time). If any stage in the chain isn't covered, Hovor blocks that run and routes you to the matching paywall before it starts, rather than silently downgrading or charging an unexpected backend.
Download Hovor and set up a Default, Casual, and Formal workflow in minutes — free tier includes 2,000 cloud words/week. Add on-device or BYOK backends anytime with Local Unlock or BYOK Unlock.
Get Hovor