One price, paid once, per number of machines.

No audio and no transcripts, ever. Audio goes from the system WebView directly to Rust over Tauri IPC, then through local whisper.cpp. There is no Electron, no Node runtime, no localhost server, no API key and no account. A packaged build makes exactly two kinds of network call, both of which you can see: a signed update check, and — only if you buy Pro — licence activation.

large-v3-turbo (1.6 GB) if you can spare the disk: it is still faster than realtime on Apple silicon, at roughly 5×, and it is substantially better on accented speech and uncommon terms. small (466 MB) is a good multilingual default. base.en (141 MB) runs at about 28× realtime and is noticeably rougher — measured here, it heard “audio workload… root the PCM” where large-v3-turbo heard “audio worklet… route the PCM”.

No, and the app never says it is. Correction memory is deterministic phrase rules plus a decoder prompt — it cannot change Whisper's acoustic model. A true fine-tune would need consented audio with reference transcripts, a training pipeline, and model conversion and serving. What the app does instead is measure: Recheck separates the model genuinely hearing more from text being rewritten afterwards, and reports both.

Whisper is measurably worse on accented speech that humans understand perfectly, and no product should promise otherwise. So 1AudioTool measures your voice rather than asserting anything about it: the benchmark scores word error rate, project-term recall and latency on your own speech, and Recheck tells you which of your terms are genuinely learned, which are being carried by a rule, and which are unstable. Explicitly selecting English rather than Auto-detect is the single biggest fix for non-native speakers on short clips.

Yes — select up to three spoken languages on the profile. Mixed speech decodes with a multilingual model, auto-detect, and a short multilingual prompt built from native-script anchors, because whisper.cpp can only lock one language token and locking the wrong one is exactly what garbles the other language. Prefer large-v3-turbo for code-switching. English-only models cannot mix, and the picker says so instead of failing quietly.

Entirely, once a model is downloaded. That includes Pro: entitlements are verified locally against a signature, with a 14-day grace window so an offline stretch or a transient failure never costs a paying customer their features.

Written, but not yet shipped. The double-Control gesture, insertion at the caret in the app you were last in, spoken responses through the system voice and the folder picker all run natively on Win32, and the release workflow can publish an NSIS installer alongside the macOS bundles. It is not in a public release yet, so this site does not offer it as a download. One caveat for when it lands: the installer is not Authenticode-signed, so SmartScreen will warn on first run.

Yes. Release the device under Settings → Licence, then activate on the new one. Profiles export as JSON — or as an AES-GCM encrypted file — and Pro can export the entire install, every profile and setting, as one file. Exports deliberately exclude microphone recordings, and never contain the licence key.

Because microphone and model lifecycle should not be coupled to an editor's process. A separate Tauri app is 3.1 MB deployed and can simply stay closed when you are not dictating; putting the same UI inside an existing Electron app would have lower incremental idle memory but would tie voice capture to that app's lifetime. The separate boundary is the cleaner one — and it exposes narrowly scoped local APIs that other applications can call.