Turn on dictation
The engine is off by default (it holds the model in RAM). Start it from the tray Voice → Start dictation, or enable auto-start on launch in the Pro Console Voice tab.
New in the Pro beta · Voice input
Voice typing for Windows, in any app — local by default.
Turn on dictation, press a hotkey, and say what you want to write — NeuroSwitcher types it into whatever app is focused: chat, browser, editor, IDE. Transcription runs on your own PC by default; or switch to zero-setup NeuroSwitcher Cloud or your own provider's API. No subscription.
- Local or cloud
- Runs on CPU
- RU · EN · DE + HE beta
- Optional AI correction
🎙 Recording… (press F8 to stop)
Let's move the meeting to Thursday at three.
Local model · transcribed on your PC
Engine: Running
Backend: Local / Cloud
Hotkey: F8
How it works
Turn it on, press F8, talk. The text lands where your cursor is.
NeuroSwitcher dictation is a small speech engine that lives next to the layout switcher you already run. It types the recognized text through the same input path used for layout corrections, so it works in ordinary text fields across Windows.
Press F8 to record
With the engine on, press the dictation hotkey (F8 by default, configurable) and speak. A small indicator — plus the Windows 11 mic icon — shows the microphone is open.
Press F8 to stop
Press the hotkey again and the microphone closes, and the text is typed at your cursor. With the local backend, audio is processed in memory — never written to disk, never sent anywhere.
Long dictation? Enable VAD
By default the whole take is transcribed when you stop. Turn on VAD segmentation (tray → Voice → VAD segmentation) to split speech at pauses: each chunk is transcribed and inserted right away — text flows phrase by phrase, and Whisper's ~30-second single-utterance cap goes away. The VAD model (Silero) is bundled, nothing extra to download.
Transcription
Choose where your speech becomes text.
Speech-to-text has three paths — pick the one that fits your privacy and setup, in the Pro Console Voice → Transcription section. Local is the default and the most private; the cloud paths trade some privacy for zero setup or frontier accuracy.
Local model · offline (default)
Download a recognition model in one click from the built-in catalog (Pro Console → Voice), with install progress. Pick a small, fast model or a larger, more accurate one. After the download, recognition runs entirely on your CPU — no internet, and no audio ever leaves your PC.
NeuroSwitcher Cloud
Zero setup: no key, no model to download, and a free daily limit. Your machine's anonymous install id signs the request. The recorded audio is sent to our proxy and forwarded to the provider. More about NeuroSwitcher Cloud →
Your own provider
Point transcription at OpenAI (e.g. gpt-4o-mini-transcribe) or any OpenAI-compatible endpoint, using your own API key. The recorded audio goes straight from your PC to that provider — no NeuroSwitcher server in between.
Switch any time
The backend is a dropdown in the Voice tab — start local for privacy, flip to Cloud when you want zero-setup accuracy, then back. Restart the voice engine to apply a change. Cloud backends are always opt-in.
When the cloud is easier: a weak PC, top accuracy without a big model download, or Hebrew without the fuss. For privacy and offline use, the local model.
Local models
Which local recognition model to pick.
For the local backend, models are downloaded from the built-in catalog (Pro Console → Voice, one-click install with progress). There is no single "best" — it depends on your languages and your PC. Our recommendations from practice:
Qwen3-ASR 0.6B · ~1 GBBest for RU/EN/DE and mixed speech — recognizes a language mix within one phrase in a single pass. The default recommendation for Russian/English/German dictation. No Hebrew.
Whisper-base · ~200 MBLight all-rounder: RU/EN/DE plus Hebrew (with the language set explicitly). A good starting point and a fit for weaker machines.
Whisper-tiny · ~110 MBThe lightest and fastest; lower accuracy. When size and speed matter more than quality.
Whisper-medium / large-v3 · ~1–2 GBTop accuracy, including Hebrew. They need more RAM and CPU — noticeably slower on a weak PC.
⚠️ Whisper detects the language once per phrase, so for speech that mixes languages within one sentence use Qwen3. For Hebrew, use Whisper (base / medium / large-v3) with the language set to HE — Qwen3 doesn't do Hebrew. The catalog also has models for other languages (Chinese, Japanese, Cantonese).
Languages
Built for people who dictate in more than one language.
Like the layout switcher itself, voice input is aimed at multilingual daily work — including sentences that switch language midway.
RussianFull dictation quality with the recommended local model.
EnglishFull dictation quality, including mixed RU/EN sentences in one breath.
GermanFull dictation quality with the recommended local model.
HebrewBeta: works through Whisper models with a language hint; recognition quality is still improving.
A recognition-language setting (Auto / RU / EN / DE / HE) is available in the tray Voice menu and in the Pro Console Voice tab. For Hebrew, setting the language explicitly is recommended.
Dictation correction
Optionally polish the raw transcript with an LLM.
Spoken text arrives with filler words, false starts, and missing punctuation. An optional LLM pass rewrites the whole dictation — punctuation, tidy filler, formatting — before it is typed. It is off by default; here is how to turn it on.
1 · Turn on Correction (LLM)Enable it in the tray Voice → Correction (LLM), or pick a Correction backend in the Pro Console Voice tab. Off by default.
2 · Pick a presetChoose a prompt preset from the tray Voice → Correction prompt — Tidy, Formal, Casual, Bullet list, Summarize, Translate, and more. The ticked preset is applied to every dictation.
3 · Edit or add promptsIn Pro Console → Voice → LLM prompt presets you can edit any built-in prompt (that becomes the new default for it) or add your own; the tray menu updates to match.
4 · Pick the AI backendFree NeuroSwitcher Cloud, your own OpenAI / OpenRouter key, or a fully local model via Ollama or LM Studio. Setup is shared with the assistant — see the AI text assistant page.
Same rules, same privacy: with a local model the correction stays on your PC; with a cloud model the transcript text goes directly from your machine to the provider you chose, using your own key (or the free NeuroSwitcher Cloud quota).
Control
Managed from the tray or the Pro Console — your pick.
Tray Voice menuStart/stop the engine, switch the active model, set the recognition language, and toggle Correction (LLM) and its preset — without opening any window.
Voice tab in Pro ConsoleModel catalog with one-click install and progress, transcription backend (local / Cloud / your key), engine status, hotkey, auto-start, and the LLM correction settings in one place.
Configurable hotkeyF8 by default; change it if it conflicts with your apps.
Opt-in engineThe voice engine only runs when you start it — it holds the model in RAM, so it stays off until you want it (or enable auto-start).
FAQ
Common questions about voice input.
How do I start dictating?
The voice engine is off by default. Turn it on from the tray Voice → Start dictation (or enable auto-start in the Pro Console Voice tab). Then press F8 to start recording, speak, and press F8 again to stop — the text is typed into the active app.
Does it work offline?
Yes, with the local transcription model — the default. After a model is installed, recognition runs entirely on your PC and needs no internet (only the one-time model download does). NeuroSwitcher Cloud and your-own-key backends are optional and do use a connection.
Is my voice recorded or stored?
With the local backend, no — the microphone is open only between an explicit start and stop, the audio is transcribed in memory, and nothing is written to disk or uploaded. If you opt into a cloud transcription backend (NeuroSwitcher Cloud or your own provider), the recorded audio is sent to that provider to be transcribed. While recording, both the NeuroSwitcher indicator and the standard Windows microphone icon are visible.
Do I need a powerful PC or a GPU?
No. Dictated phrases are short, and the default local models transcribe them faster than real time on an ordinary CPU. Larger, more accurate models are available in the catalog if your machine can handle them — or offload to a cloud backend entirely.
Which apps does it work in?
Any app with a normal text field: browsers, chats, editors, IDEs, office apps. The text is typed through the same input path NeuroSwitcher uses for layout corrections, and the same per-app exclusion list applies.
How good is Hebrew?
Hebrew dictation is in beta. It works through Whisper models with the language explicitly set to Hebrew, and the output is real Hebrew text, but recognition quality is noticeably behind Russian, English, and German for now. This is being actively worked on.
Is voice input free?
Voice input is part of the free Pro beta. The managed model catalog with one-click install is a Pro convenience; the underlying voice engine is part of the open core. NeuroSwitcher Cloud has a free daily limit; your-own-key backends bill through your provider.
My antivirus asks for microphone access for “ks-voice-host.exe” — is that expected?
Yes. The microphone is opened by a separate recognition process, ks-voice-host.exe, in the voice subfolder of the NeuroSwitcher install — not by the main application. Recognition runs out of process on purpose: the speech model stays isolated from the layout switcher, so a problem in one cannot take the other down. That is why the prompt from Windows or your security software names a binary you have not seen before instead of “NeuroSwitcher”. Allowing it once is enough.
Free Pro beta
Type with your voice — your way.
Download NeuroSwitcher for Windows 10 and 11, open the Voice tab, turn on dictation, and speak into any app — with a local model offline, or NeuroSwitcher Cloud with nothing to set up.
Download NeuroSwitcher