Wispr Flow popularized cloud voice typing — but its trade-offs (audio routed through OpenAI and Meta, $144/year subscription, no on-device option) push privacy-conscious users toward something different. Hovor is built on the opposite premise: voice typing that runs on your iPhone and Mac, not on someone else's server.
| Feature | Wispr Flow | Hovor |
|---|---|---|
| On-device transcription | No (cloud only) | Yes — Parakeet v3 + Whisper |
| iPhone keyboard | Yes | Yes (system-wide) |
| macOS native | Yes (Electron) | Yes (native SwiftUI) |
| Windows | Yes (Electron) | In development — native, coming soon |
| Ukrainian support | Listed; not "pro-tuned" | Strong via Parakeet v3 (NVIDIA, 24 EU languages) |
| Slavic languages | Listed | Parakeet v3 covers PL, CS, SK, BG, HR, SL, SR |
| Recording length | 6 min cap | Unlimited |
| BYOK (bring your own key) | No | Yes — OpenAI, Anthropic, custom endpoints |
| Audio storage policy | Routed via OpenAI/Meta | Discarded immediately, never stored |
| Subscription | $12–15/mo, $144/yr | $11.99/mo, $89.99/yr (with on-device free tier) |
Wispr Flow has real strengths. Its AI auto-edit cleans up filler words and run-on sentences in a way that feels almost magical the first time. The marketing budget (post-$81M raise) means polished onboarding, frequent updates, and a brand presence that none of its competitors match.
But the design choices that enable that magic are also its weaknesses for a particular kind of user:
Hovor ships with two on-device speech recognition models: Parakeet v3 (NVIDIA, 0.6B parameters, supports 24 European languages including Ukrainian) and Whisper Large V3 Turbo (4-bit quantized for iPhone). On modern iPhones with the Neural Engine, they run far faster than real-time — six minutes of audio transcribed in 1.7 seconds on iPhone 15 Pro Max.
Cloud dictation via OpenAI is available too, but it's an option, not a requirement. The privacy story isn't marketing — it's the default code path.
To be clear: we don't train the speech models. Parakeet v3 is built by NVIDIA and released under CC-BY-4.0; Whisper is OpenAI's open-source release; EuroLLM (used for the cleanup pass on Mac) is Unbabel's. Our job is to pick the open models that work best for our users and wire them into the app properly.
For Ukrainian and Slavic users, that choice is significant. Wispr Flow's "100+ languages" claim covers quality that ranges from excellent (English, Spanish, French) to mediocre for most Slavic languages. Parakeet v3 was trained on 24 European languages — Ukrainian, Polish, Czech, Slovak, Bulgarian, Croatian, Slovenian, Serbian (Cyrillic), and others — with explicit balancing across them. For users who switch between languages mid-sentence, the underlying model architecture matters more than benchmark headlines.
Where Hovor's own engineering shows up is in the surrounding pipeline: prompt design that explicitly forbids unsolicited translation (a common failure mode for Ukrainian text passed through gpt-4o-mini), a custom dictionary for proper nouns the model routinely mishears, and a UI fully translated into Ukrainian.
Hovor lets you use your own OpenAI, Anthropic, or custom OpenAI-compatible endpoint (Ollama, LM Studio, vLLM) API key for the formatting/cleanup step. You pay the API provider directly — Hovor never proxies your key, and the key never leaves your device's Keychain. This is the opposite of Wispr Flow, which doesn't support BYOK at all.
Why does BYOK matter? In 2026, most professional users already have API access through their work or personal subscription. Paying $144/year for Wispr's UI when you already pay OpenAI for the model output is a tax on infrastructure you don't need.
Hovor for Mac is a SwiftUI app. Hovor for iOS is native SwiftUI with a custom system keyboard extension. The Windows client is being rebuilt natively (Rust + React via Tauri) and is coming soon — no Chromium runtime, no 800 MB RAM baseline, no foreground-app freezes during paste. Cold start is under 1 second on every shipping platform.
No 6-minute cap. If you want to dictate a 30-minute draft, hit record and talk. The recording state machine handles long sessions natively because Parakeet v3 is built for streaming inference up to 24 minutes per inference window.
Hovor's dictionary supports multi-replacement variants (v1|v2|v3 → canonical form) — useful for proper nouns, technical jargon, or names that voice models routinely mishear. Snippets work in any text field, with private snippets marked local-only (they never sync, never leave the device). Wispr has the dictionary feature, but the implementation is simpler.
Honest section, because there's no point pretending otherwise:
For everyone else — and especially for Ukrainian speakers, Slavic-language users, or anyone who doesn't want their voice routed through three companies' servers — Hovor is worth a look.
| Wispr Flow | Hovor | |
|---|---|---|
| Free tier | 2,000 words/week (cloud) | 2,000 words/week cloud + on-device free weekly quota |
| Monthly | $12–15 | $11.99 |
| Yearly | $144 | $89.99 (~$7.50/mo) |
| Lifetime / one-time | None | Local Unlock $49.99 · BYOK Unlock $24.99 · Bundle $69.99 |
| BYOK cost | Not supported | $0 — pay your provider directly |
| Regional pricing for Ukraine | Standard USD | ~50% reduction (UAH) |
Privacy is overused as a marketing word. Here's what "privacy-by-default" looks like in Hovor's actual architecture:
Input Monitoring or screen-content permissions. The keyboard extension only reads the text field it's typing into — same scope as Apple's stock keyboard."On-device by default" is not a slogan — it's a code path. The cloud is opt-in, not the foundation.
For on-device transcription on Apple Silicon: yes, by a large margin. Wispr's cloud round-trip is 500ms+ per request (network + processing + network back). Hovor's on-device path is local inference only — typically 30–50ms per word streamed back. For comparable cloud paths (both using OpenAI Whisper), latencies are similar.
Yes, with on-device models. Wispr Flow does not (its recently added "offline fallback" is still cloud-dependent for the cleanup pass).
Apple Intelligence is system-level and works for general dictation, but it doesn't support a custom dictionary, doesn't offer per-app modes, and has limited Ukrainian quality. Hovor coexists with it — many users keep both, depending on the context.
For HIPAA-grade compliance, you'd run Hovor entirely on-device (no cloud STT, no cloud format, BYOK only if needed for AI cleanup, with a self-hosted endpoint). The architecture supports this; we don't currently offer a HIPAA BAA, so consult your compliance team for the specific certification you need.
The Free tier is permanent, not a trial. It includes 2,000 words/week of cloud STT, a free weekly quota for on-device transcription (unlimited with Local Unlock), and the system keyboard. Pro removes the cloud quota and adds 5-device sync, custom workflows, and BYOK across all providers.
iPhone and Mac today, Windows coming soon. Free tier, no credit card required. On-device by default. Quality Ukrainian support.
Get HovorLooking for Superwhisper alternative? · Voice typing in Ukrainian · Українською