Free offline Cantonese speech-to-text for Windows. Speak naturally in Cantonese and your words are instantly converted to text — no internet required, fully private local inference.
Type4All is a free Cantonese voice-input application for Windows. Hold a hotkey, speak in Cantonese (or mixed Cantonese + English), and the spoken text is typed directly into whatever application has focus — Word, Chrome, WhatsApp Web, Notepad, comment boxes, anywhere a normal keyboard would type.
It is built specifically for the Hong Kong context: it understands HK colloquial vocabulary, code-switched English terms, and proper nouns commonly used locally. It runs entirely on your own computer after a one-time AI-model download, so once installed there is no cloud round-trip and no recurring subscription.
Type4All is part of the AI4All initiative — a project aimed at making practical AI tools accessible to ordinary users in Hong Kong, so that the benefits of AI are not limited to people fluent in English or comfortable with technical jargon.
Type4All bundles a Cantonese-capable speech-recognition (ASR) model that runs locally on your machine using ONNX Runtime. The high-level flow is:
1. You hold a configured global hotkey (default: Right Ctrl).
2. While the hotkey is held, the microphone records audio.
3. When you release, the audio is passed to the local ASR model for transcription.
4. The recognised text is typed into the active text field via simulated keystrokes — exactly as if you had typed it.
Key architectural choices:
• Local-first: After a one-time ~3 GB model download on first run, all transcription happens on your CPU. No audio is sent to any server.
• Offline-capable: Once the model is downloaded, the app does not require internet to function.
• Universal compatibility: Output is plain keystrokes, so Type4All works in any application that accepts keyboard input — no per-app integration needed.
• Lightweight footprint: The keyboard hook only checks whether the pressed key matches your configured hotkey; it does not log other keystrokes.
Because inference runs locally, real-world latency depends on your hardware. SSD + 16 GB RAM produces near-instant transcription; older HDD-only machines will be noticeably slower.
Two reasons.
First, the original mission. Type4All exists because Windows has historically had weak Cantonese voice-input support, and most third-party solutions are either cloud-based (raising privacy concerns) or paid. A Hong Kong user who simply wants to dictate in Cantonese without sending their voice to a foreign cloud, and without a monthly fee, had limited options. Charging for Type4All would defeat the point of solving that gap.
Second, sustainability comes from elsewhere. The AI4All YouTube channel (https://www.youtube.com/@AI4All-edu) shares Cantonese AI tutorials freely, and is considering optional paid offerings — for example, advanced courses for users who want to go deeper, or channel membership for community-focused content. Those optional revenue streams are what allow free tools like Type4All, and free educational content, to keep being made.
In short: free is the product, not a teaser. There is no plan to introduce a paid "pro" tier of Type4All itself, no upsell screens, and no telemetry collected to monetise users in any other way.
Yes. Type4All is a free Cantonese voice typing app for Windows. Press a hotkey, speak in Cantonese, and the recognised text is typed straight into whatever app is in front of you. After a one-time AI model download, it runs fully offline — no subscription, no usage limits.
Yes — Type4All. It runs a local AI speech model entirely on your PC, so your voice never leaves your computer. After a one-time AI model download during first-run setup, no internet connection is needed for dictation. No cloud upload, no telemetry.
Yes. Type4All works in any Windows app that accepts text input — including WhatsApp Desktop, Microsoft Word, Notion, Outlook, browsers, and most chat or note-taking tools. Press your global hotkey to start dictation; the recognised Cantonese text is automatically pasted into wherever your cursor is. No per-app integration or plugin needed.
That's the first-run terms / privacy acknowledgement screen introduced in v0.8.6.
Because Type4All uses your microphone and a global keyboard hotkey, we want every user to see the following in plain language before they start:
• The microphone is only active while you hold your configured hotkey.
• The keyboard hook only checks whether the pressed key is your hotkey — it does not log any other keystrokes.
• After the one-time AI-model download, the app runs entirely offline and sends no telemetry.
• The software is provided "AS IS" without warranty of any kind.
Once you click "Accept and continue", the app remembers your choice and won't show the screen again.
Full details:
• Privacy Notice Section 5A — https://4all.dev/en/privacy
• Terms of Service, "Downloadable Software Applications" section — https://4all.dev/en/terms
• DISCLAIMER.txt — shipped with the application, in the unpacked Type4All folder
This warning means Windows microphone permission hasn't been granted yet. Try the steps below.
Fastest method:
Press Win + R, type ms-settings:privacy-microphone, and press Enter — this jumps straight to the microphone permission page.
Or manually: Start → Settings → Privacy & security → Microphone
Make sure all three toggles are ON:
1. Microphone access
2. Let apps access your microphone
3. Let desktop apps access your microphone ← most commonly missed
Pay special attention to the third toggle — Type4All is a desktop app, managed separately from apps downloaded from the Microsoft Store.
After that: close Type4All and reopen it.
First launch loads a ~3 GB AI speech model. Typical timings:
• SSD + 16 GB RAM (recommended): first launch 10–30 seconds / subsequent 5–15 seconds
• SSD + 8 GB RAM: first launch 30 seconds–1 minute / subsequent 10–30 seconds
• HDD (traditional spinning drive): first launch 3–15 minutes / subsequent 1–5 minutes
If you've waited over 2 minutes and it's still not ready, the most likely causes are:
• Your PC is using a traditional HDD (not an SSD)
• Or available RAM is too low (the model needs around 3 GB of contiguous memory to load)
We're sorry — the AI model has certain hardware requirements; an SSD is needed for smooth operation.
Reason: Type4All runs an AI language model locally on your PC, which requires adequate memory and CPU.
Steps to try (in order):
1. Close Type4All → restart your PC → open only Type4All and wait 1–2 minutes for cold start
2. Open Task Manager and check "Available memory" — keep at least 4GB free
3. Make sure you have the latest version (v0.9.1 or later)
If none of the above work, see Q6 below to export a debug log and email it to us.
Same root cause as Q4 — usually insufficient RAM during model inference causes the model process to be terminated by the system.
Suggestion: Close other heavy applications (Chrome with many tabs, games, video editors, etc.) and try again.
If the problem persists, follow Q6 below to export a debug log and send it to us.
v0.7.8 added a "Debug Log" feature so you can review it yourself or share it with us for debugging.
How to open it?
1. Right-click the floating microphone panel → select [Show Debug Log]
2. The small window shows in real time: system info (CPU, RAM, Python / ONNX versions), recognition time per utterance, error messages
Important: We never record what you say — only metadata (timestamps, duration, error messages). Safe to share.
Report a problem?
Click [Save] to export the .log file → email to [email protected], and we'll take a look.
━━━━━━━━━━━━━━━━━━━━━━━━━━
💡 Friendly note
Type4All's AI model runs entirely locally on your machine (truly offline, fully private),
so real-world performance depends on your hardware and Windows configuration.
Some machines may have compatibility issues,
and certain hybrid-architecture CPUs (especially three-tier P-core + E-core designs)
may not run smoothly on every system. We appreciate your understanding 🙏
━━━━━━━━━━━━━━━━━━━━━━━━━━
• OS: Windows 10 / 11 (64-bit)
• RAM: 16 GB minimum, 32 GB recommended (needs at least 8 GB free when idle)
• CPU: Any mainstream Intel / AMD CPU from the last 5 years (older Atom / Celeron not recommended)
• Disk: 5.5 GB+ free space
• Network: Required only for initial model download; fully offline afterwards
It isn't a virus — this is a standard warning shown by Windows SmartScreen and some antivirus tools against "unsigned" software.
Why it appears:
• Type4All is a small free tool from an independent developer, licensed under the Apache License 2.0. We haven't purchased a code-signing certificate yet (about £100–300/yr), so Windows doesn't yet "recognise" the program.
• Type4All uses a global keyboard hook to detect your configured hotkey. This technique is also used by some malicious software, so a small number of AV tools raise a false positive.
How to continue the install:
1. On the SmartScreen warning window, click "More info" → "Run anyway".
2. If Defender has already quarantined the file, go to "Virus & threat protection → Protection history → Allow" to add Type4All.exe to the allow list.
Why you can trust it:
• After the one-time AI-model download, the app runs entirely offline and sends no data, crash reports, analytics, or telemetry anywhere.
• The global keyboard hook only checks whether the pressed key is your configured hotkey — it doesn't log any other keystrokes.
• Full details: Privacy Notice Section 5A (https://4all.dev/en/privacy) and the DISCLAIMER.txt shipped with the application.
The most common cause is that Right Ctrl has already been claimed by another program, so Type4All cannot register it as a global hotkey. Common sources of this conflict include:
• Chinese input methods: Sogou, Microsoft New Phonetic, Baidu IME and similar tools may have assigned Right Ctrl as their input-method toggle key.
• Hotkey manager software: AutoHotkey, PowerToys, Logitech Options, Razer Synapse and similar tools may have already registered Right Ctrl as a global hotkey.
In addition, Windows itself has inherent limitations on using "pure modifier keys" as global hotkeys — modifier keys (Ctrl, Alt, Shift) were originally designed to be combined with other keys. When registered alone, they tend to conflict with the system, input methods, or other software.
If the shortcut is temporarily unavailable, you can use the mouse to click the recording button on the Type4All window as an alternative.
We are considering adding combo-key support (for example, Ctrl + Alt) in the next version to reduce the likelihood of conflicts.