Frequently asked questions

On first launch the app detects your CPU, RAM and GPU and shows a "💡 Recommended for your hardware: …" badge in settings. Weak laptop → Whisper Base. Mid → Small. Strong with CUDA → Large-v3. No manual trial-and-error. No discrete GPU? Still works — just on CPU.

Yes. On first launch it reads your Windows system locale and picks the right model automatically: en_US → English with the Zipformer model, de_DE → German with Whisper, ja_JP → Japanese with Whisper, etc. The UI itself ships in your language (RU/EN/DE/ES/FR/IT). No setup needed.

Windows 10 (build 1809+) or 11. 4 GB RAM for the base model. ~500 MB disk. Any microphone. NVIDIA GPU optional for speedup.

No. Minimum is Windows 10. This is due to modern speech recognition libraries.

Planned, no exact date. Subscribe — we will tell you.

Right now — installer only (Inno Setup), no portable version. We're considering a portable build for 1.0, but it requires admin rights anyway for tray hotkey hooks. If you can't install programs at work — sorry, AuroraWhisp won't help you yet.

Really. Recognition runs through local models. You can verify with Wireshark or block internet for the app via firewall — it will keep working.

Updates are the only thing the app downloads from the internet. And you can disable that — then the app never goes anywhere.

No. Never. We use pretrained open recognition models that run locally. Your voice does not leave your computer, and we physically cannot get it.

$19.90. One-time, forever.

Unlimited words (no daily 5,000 cap), access to heavy Whisper Medium / Large-v3 / Distil-Large-v3 models (GPU recommended for comfort), all 12 widget styles, unlimited custom replacement rules, priority email support, license for 3 devices, no-questions 14-day refund. All future updates — free.

Write to hello@aurorawhisp.com — we will issue a licence and send a payment link. Automatic online checkout coming.

The program keeps working offline as long as Windows works. Pro features will not turn off because the license is verified locally.

It will. The license is for life. The price you pay now is locked in for you. All future updates — free.

All 15 — with verified WER<15% quality: English, Spanish, German, French, Italian, Portuguese, Dutch, Polish, Czech, Turkish, Russian, Ukrainian, Japanese, Korean, Chinese. For English we ship a dedicated fast model (Sherpa Zipformer) — ~150 ms on a 5-second phrase. The other languages run on Whisper at comparable quality.

Technically the model can recognize many languages, but we only guarantee quality on the main 10. Try it — most likely it works.

Not yet. One dictation session = one language. Switching is one click in settings. Auto-detect mid-dictation is on the to-do list.

Ctrl+Space by default. Many also pick F9, right Ctrl, or Caps Lock. Depends on your keyboard and other apps. If Ctrl+Space is taken (e.g., for autocomplete in your IDE) — use F9, also great.

Settings → "Replacements" → add by voice: say the word 5–6 times with different intonation, the app remembers all variants.

Depends on the game. In most — yes (text inserted in chat). In some exclusive-fullscreen games Windows blocks input from other apps.

Zipformer English is a streaming model from k2-fsa, tuned specifically for low-latency CPU inference. Unlike Whisper, which processes the full clip after you finish, Zipformer starts recognising while you are still speaking and holding the key. By release most of the phrase is already done — hence ~150 ms on a 5-second phrase vs ~1 second for Whisper Small on the same CPU.

GPU is only needed for heavy Whisper models: Medium / Large-v3 / Distil-Large-v3. They run on CPU, but slowly (3-7 sec on a 5-sec phrase). With NVIDIA + CUDA — 100-300 ms. Sherpa Zipformer English runs on CPU at ~150 ms — no GPU needed. On a laptop without a GPU pick Sherpa Zipformer or Whisper Tiny/Base/Small and you will be fine.

Idle (in tray, not dictating) — almost zero: ~30 MB RAM, 0% CPU. While dictating — short bursts of ~10-30% on one core for ~200 ms. Across a normal day (an hour of dictation total) AuroraWhisp drains less than 1% of battery. Heavy Whisper Medium/Large on CPU can heat up the laptop noticeably — on a laptop pick Sherpa models or Tiny/Base.

Baseline ~30 MB when no model is loaded. After the first dictation the model stays in memory: Zipformer English — ~120 MB, Whisper Tiny — ~80 MB, Base — ~150 MB, Small — ~500 MB, Medium — ~1.5 GB, Large-v3 — ~3 GB. On a system with 8 GB RAM, Small is comfortable. With 16 GB — Medium/Large is fine.

RTF (Real Time Factor) is the ratio of recognition time to audio length. RTF 0.03 for Zipformer English on CPU means: 1 second of speech is recognised in about 30 milliseconds — over 30× faster than real-time. For comparison: Whisper Medium on CPU has RTF ~1.0-1.5 (slower than or equal to real time). Hence Sherpa Zipformer is not just "fast" but "you can transcribe long files in seconds in the background".

Auto-update is the **only** regular internet connection (plus a one-off Pro activation on first install). Every 24 hours the app makes an HTTPS request to aurorawhisp.com/api/updates/latest.json — gets a JSON with the version and SHA-256 of the new exe. If there is a new version — it asks you whether to download. You can disable it entirely in Settings → Updates → "Automatically check for updates: off". After that the app never goes online.

A false positive because the exe is not signed with a code-signing cert — not bought yet. Fix: open antivirus settings → Exclusions / Trusted apps → add AuroraWhisp.exe. After that the antivirus stops reacting. If it does not work — write to bugs@aurorawhisp.com, we will help for your specific antivirus.

Wispr is cloud-based, $15/mo, AI-cleans your speech. We are local, $19.90 once, no text editing. Need automatic polish — Wispr; need privacy and no subscription — us. Full comparison: /en/compare/wispr-flow.

WhispeRu is Russian only, 4,990 ₽ once. We are 15 languages, 1,490 ₽ / $19.90 once (a third of the price). Both local. If only Russian — pick by tone. If you need more — we have the full pack. Full: /en/compare/whisperu.

Win+H is cloud-based (Microsoft Cloud), needs an MS account, struggles with long phrases and rare words. We are local, no account, better on long phrases and custom terms. Full: /en/compare/windows-voice-typing.

Dragon is an old $300+ product tuned for doctors and lawyers with specialised vocabularies. We are modern at $19.90, for everyone else. If you are not dictating medical records or legal briefs — Dragon is not for you. Full: /en/compare/dragon.