Documentation

Get up and running in 60 seconds

Install OFFTYPE, press a hotkey, start speaking. This page covers everything else.

What is OFFTYPE

OFFTYPE is a fully offline voice input tool. Speak naturally, and your words appear as text in any application. No internet required. Your data never leaves your computer.

  • Voice Input: Speak → text appears where you’re typing
  • Translate: Speak Chinese → output in English, Japanese, Korean, French, Spanish, or German
  • Polish: Rewrite casual speech into formal, casual, or simplified styles
  • Custom Dictionary: Define how names, terms, and abbreviations should be written

System Requirements

항목최소 사양
OSWindows 10 64-bit or later
RAM8 GB
Disk2 GB (including SenseVoice model)
NetworkOnly needed for initial model download

Installation

  1. Double-click OFFTYPE-Setup-1.0.0.exe

  2. Follow the wizard (choose install location, desktop shortcut)

  3. OFFTYPE launches automatically when done

Security notice: Windows Controlled Folder Access

OFFTYPE embeds llama.cpp (MIT open source license) for on-device translation and text polishing. The Windows executable for this open source project is called llama-cli.exe. Because it lacks a commercial code signing certificate ($200–400/year), Windows Security may flag it as an unrecognized application.

If you see “Controlled folder access blocked llama-cli.exe”:

  1. Open Windows Security → Virus & threat protection
  2. Under "Controlled folder access", click "Manage Controlled folder access"
  3. Click "Allow an app through Controlled folder access" → Add an allowed app
  4. Browse to C:\Program Files\OFFTYPE\speech-server\_internal\llama-cpp\llama-cli.exe
  5. Confirm

What happens if I skip this? Translation and polish won’t work. Voice input (ASR) is unaffected. It’s your choice.

Why isn't this done automatically? Modifying system security settings is a significant change to your computer. OFFTYPE won’t make that decision for you.

First Launch

OFFTYPE connects to its local service on startup. If models aren't downloaded yet, you'll be prompted.

Required model

ModelSizePurpose
SenseVoice~239 MBSpeech recognition (mandatory)

A download window opens automatically on first launch.

Optional models

ModelSizePurpose
Qwen3-ASR~988 MBWider coverage: 30 languages + 22 Chinese dialects (Cantonese, Sichuanese, Vietnamese, Arabic, etc.)
Local LLM (1.5B)~1.04 GBOn-device translation & polish. No Ollama needed

Voice input works without the Local LLM. For translation and polish, you can also install Ollama and use larger models.

Basic Usage

Push to Talk (PTT)

  1. Click "Push to Talk"
  2. Hold Ctrl+Space and start speaking
  3. Release — text appears at your cursor

Best for short sentences. This is the default mode.

Voice Activity Detection (VAD)

  1. Click to expand the settings panel
  2. Set Trigger to VAD
  3. Start speaking — OFFTYPE detects sentences automatically and types each one when you pause

No keys needed. Best for longer dictation.

How to tell if VAD is working?

Watch the Push to Talk button color:

  • Pulsing green (breathing light): Speech detected, working normally
  • Solid green, never blinks: VAD engine is running but not detecting speech. Try restarting OFFTYPE. If the light stays solid, check that the ASR model is downloaded in Preferences

Switching Modes

Use the Process dropdown in the expanded panel:

ModeResult
InputSpeak → text (default)
TranslateSpeak Chinese → translated text
PolishSpeak → rewritten in your chosen style

Settings Panel

Click to expand. Four options:

  • Trigger: PTT (push-to-talk) or VAD (continuous listening)
  • Process: Input / Translate / Polish
  • Target: Output language (translate mode)
  • Style: Polish style (Formal / Casual / Simple)

Preferences

Click the gear icon to open.

ASR Model

Speech recognition model. Downloaded models appear in the list; missing ones show “(download required).”

  • SenseVoice: Default. 5 languages (Chinese, English, Japanese, Korean, Cantonese). Fast.
  • Qwen3-ASR: 30 languages + 22 Chinese dialects (Sichuanese, Vietnamese, Arabic, Hindi, etc.). Slower than SenseVoice.

Ollama

If you’ve installed Ollama (ollama pull qwen2.5:7b), enable this to use a 7B model for translation and polish. Better quality than the local 1.5B model.

Local LLM

Qwen2.5 1.5B on-device model. Powers translation and polish when Ollama is off. Click Download to install.

Dictionary

Custom word replacement editor. Add fixed-term mappings:

Spoken (ASR output)Replace with
off typeOFFTYPE
alley babaAlibaba

Dictionary is plain text replacement — it cannot fix accent-related errors (e.g. “four” vs “for”). Use Qwen3-ASR for dialect issues.

Fast Typing

Uses SenseVoice's built-in punctuation and skips LLM correction. Faster, but punctuation may be less accurate. Good for quick input.

System Tray

Close the window (X) → minimizes to tray. OFFTYPE keeps running in the background. Double-click the tray icon to restore, or right-click:

  • Show: Restore window
  • Exit: Quit completely

Troubleshooting

Translation / polish not working?

  1. Is Ollama enabled in Preferences but Ollama isn’t running? (OFFTYPE will wait for the connection and time out)
  2. Is the Local LLM downloaded?
  3. Is Windows Controlled Folder Access blocking llama-cli.exe? (See Installation section above)

Download is slow?

  • SenseVoice downloads from HuggingFace / ModelScope. In most regions, ModelScope picks up automatically.
  • Qwen3-ASR downloads from GitHub Releases. OFFTYPE probes available mirrors and uses the fastest one.

Nothing happens after launch?

  1. Wait 5 seconds — the server needs time to start
  2. Open Task Manager to check if speech-server.exe is running
  3. If a previous version was installed, kill old processes first: taskkill /F /IM speech-server.exe

How do I uninstall?

  • Start Menu → Settings → Apps → Installed apps → Search OFFTYPE → Uninstall.
  • Model files are stored in %APPDATA%\OFFTYPE\models\. You’ll need to delete this folder manually after uninstalling.

FAQ

Does OFFTYPE send my voice anywhere?

No. Voice recognition, punctuation, and translation all happen on your machine. See Privacy for the Wireshark proof.

Does OFFTYPE work offline?

Yes, after the first model download. Subsequent use is fully offline.

Can I use it on macOS or Linux?

Not currently. OFFTYPE is Windows-only. macOS and Linux are not on the immediate roadmap.

Is there a paid version?

OFFTYPE is currently free. There is no subscription, no account, no payment.

Can I run it on a corporate network?

Yes. OFFTYPE only needs network access for the initial model download and optional update checks. After that, it works fully offline. If your firewall blocks Hugging Face, configure the ModelScope mirror.

How do I uninstall?

Windows Settings → Apps → OFFTYPE → Uninstall. To remove all data, delete `%APPDATA%\OFFTYPE\` manually.

Getting support

If something doesn't work, please email us with:

  • Your Windows version
  • OFFTYPE version (tray icon → About)
  • What you were trying to do
  • Any error message you see

Still stuck? Email support@offtype.app with your Windows version, OFFTYPE version, what you were trying to do, and any error message. We read every message.