SpeakTypeFully open source
MIT licensed. Protocols, correction algorithms, UI — every line is readable, hackable and self-hostable.
v0.17.2 Windows stable · macOS preview · Doubao ready right after sign-in
You speak,
You speak,
it types.
Hold RightCtrl, talk, release — your words land at the cursor in any app. Recognition runs on your machine by default; engines, AI polishing and the hotword dictionary are all yours to configure.
- MIT licensed
- Offline by default
- No account, no telemetry
- 5 UI languages
Windows 10/11 x64 · macOS preview (Apple Silicon) · release notes & older builds
0servers of its own — audio stays local unless you pick a cloud engine
4offline model options: SenseVoice, Parakeet int8 / fp32, Whisper
5UI languages: English, 简体中文, 繁體中文, 日本語, 한국어
MITopen source — protocols, correction logic and UI all readable
Why SpeakType
Dictation that belongs to you
Most AI dictation apps are closed source and route your voice through the vendor's servers. SpeakType has none — audio goes only where you tell it to go, or nowhere at all.
Local-first, no backend
Recognition happens on your machine by default. Audio leaves it only if you configure a cloud engine — and then only to that endpoint.
Learns from your edits
Fix a wrong word by hand after it lands and SpeakType learns {wrong → right} into your dictionary. Same mistake, never twice.
Hands-free mode
Tap Alt+Q or double-tap the hold key. Sentences split on silence, longer pauses start a paragraph, Esc cancels anytime. Optional voice commands.
Select & rewrite
Select text, hold F8 and say “translate to English” or “make it formal” — the selection is replaced in place. Switch windows meanwhile and the result waits in your clipboard.
Personas per app
Alt+1…9 switches styles — default, auto-translate, report-to-boss, CLI, your own prompt — or let them switch automatically by foreground app.
Your phone as the mic
No microphone on the desktop? Scan a QR code and talk into your phone — LAN direct, or through a relay you can self-host.
File transcription
Drop audio or video (mp3, wav, m4a, mp4… up to 3 h). Offline segmented transcript with timestamps; export TXT, SRT or VTT.
Sixty seconds
From download to dictating
Install
Run the installer, or unzip the portable build — it keeps its config next to the .exe.
Pick an engine
Settings → Speech → built-in offline model (one-click download), or paste your own API key.
Hold and talk
Put the cursor anywhere, hold RightCtrl (or a mouse side button), speak, release.
It learns
Fix a word by hand and it goes into your dictionary automatically — accuracy compounds.
Recognition
Four offline models, or bring your own key
Download a model once inside the app and stay offline forever — or point SpeakType at any OpenAI-compatible endpoint. Switch anytime.
| Model | Size | Best for | Notes |
|---|---|---|---|
| SenseVoice Smalldefault | 234 MB | Chinese, Cantonese, Japanese, Korean, English | ~0.27 s per utterance in our tests, punctuation built in |
| Parakeet TDT 0.6B v3int8 | 660 MB | English + 25 European languages | Highest English accuracy; int8 occasionally clips the first word of a sentence |
| Parakeet full precisionfp32 · optional | 2.5 GB | Same languages as Parakeet | Fixes the clipped first word; ~2.7 GB RAM while loaded, same speed as int8 in our measurements |
| Whisper tiny / base / smallwhisper.cpp · Windows only | 32 / 60 / 190 MB | Broadest language coverage | tiny is fastest but error-prone; small is slowest and most accurate |
Swipe sideways to see the whole table
Resumes after interruption
SHA-256 verified
3 sources: Hugging Face → hf-mirror → GitHub Releases
Switches source when a transfer stalls
Plain-language errors with retry
Any OpenAI-compatible API
- Base URL + API key + model name for
/audio/transcriptions - Presets: OpenAI Whisper, Groq (free tier), Fireworks, Mistral Voxtral, SiliconFlow, Alibaba Bailian
- One-click connection test
Web providers you sign into
- ChatGPT web transcription (a free OpenAI account works)
- Doubao voice — ready right after sign-in
- Off by default; undocumented endpoints, read the disclaimer
Any chat model, or none
- OpenAI, Gemini, Groq, DeepSeek, Zhipu GLM-4-Flash, Kimi, Qwen…
- Local Ollama / LM Studio endpoints (API key can stay empty)
- Without one, a local pass still fixes self-corrections (“5pm — no, 6pm” → “6pm”)
A closer look
Built for daily use, light or dark
Follows your OS theme in real time. Everything below is a real screenshot of the Windows build.
Phone as microphone
No mic on the desktop? Use the one in your pocket.
Scan the QR code in Settings, hold the button on your phone and talk. The text lands on the computer, at the cursor, exactly like a local recording.
- LAN direct — phone and PC on the same Wi-Fi, nothing leaves your network.
- Relay when you're apart — the official relay is a tiny Cloudflare Worker you can self-host in one command.
- Browser or app — open the link in any phone browser, or install the small Android APK.
- Robust pairing — reconnects after a refresh, hands over cleanly when a second phone takes the room.
Compare
How it stacks up
Against the closed-source dictation tools people usually reach for.
| SpeakType | Typical AI dictation app | OS built-in dictation | |
|---|---|---|---|
| Source code | Open (MIT) | Closed | Closed |
| Works fully offline | Yes — 4 local models | Usually cloud-only | Partly |
| Where your audio goes | Nowhere, or the endpoint you choose | Vendor servers | Vendor servers |
| Bring your own engine / key | Yes | No | No |
| Learns corrections from your edits | Automatic | Manual dictionary | No |
| Hands-free & rewrite selection | Yes | Varies | Limited |
| Phone as microphone | Yes | No | No |
| Price | Free | Subscription | Free |
Swipe sideways to see the whole table
Download
Get SpeakType
Windows v0.17.2 is the stable release. The macOS build is an early preview — please read the notes before installing.
Windows
Stable · v0.17.2scoop bucket add speaktype https://github.com/wookat/scoop-speaktype
scoop install speaktypeUnsigned installer — if SmartScreen objects, choose “More info → Run anyway”.
macOS
PreviewNot notarized: allow it under Privacy & Security or run xattr -d com.apple.quarantine /Applications/SpeakType.app. Hold Right Option to talk. Local engines: SenseVoice and Parakeet.
Android
Phone micNot a standalone dictation app: it streams your voice to SpeakType on the desktop over LAN or the relay.
What we have verified
- Every Windows build on a real Windows 11 machine: packaged installer, simulated microphone and hotkeys, 100 / 125 / 150 % DPI
- macOS preview on an Apple Silicon machine: build, install, permissions, hold-to-talk → text lands in TextEdit
- Parakeet fp32 on an offline A/B corpus and a real 2.5 GB download
- Phone microphone over LAN and the public relay (browser-simulated phone), including takeover and reconnect
What we have not — stated plainly
- macOS: signing / notarization, Intel hardware, physical microphones and keyboards; “learn from my edits” and whisper.cpp are Windows-only
- Windows: physical microphone hardware (tests use a simulated audio device) and cloud providers with real paid keys
- A physical phone or the Android APK on a real device; long-term fp32 memory behaviour on low-RAM machines
FAQ
Questions, answered
Does my voice leave my computer?
Not unless you choose a cloud engine. With the built-in offline models, recognition, punctuation and the “learn from my edits” comparison all run on your machine. SpeakType operates no servers and collects nothing; API keys and history live in %APPDATA%\SpeakType (Windows) or ~/Library/Application Support/SpeakType (macOS).
Which languages does it recognize?
SenseVoice covers Chinese, Cantonese, Japanese, Korean and English; Parakeet covers English plus 25 European languages; Whisper (Windows) covers the broadest set. Cloud engines follow their own language lists. The UI itself is available in English, Simplified and Traditional Chinese, Japanese and Korean.
Why is the installer unsigned?
Code-signing certificates cost money the project does not have yet. The installer is built from the public source; SmartScreen shows “More info → Run anyway”. On macOS the app is ad-hoc signed and not notarized — see the install guide for the one-line xattr workaround.
Is the macOS version ready for daily use?
It is a preview. Building, installing, granting permissions and hold-to-talk dictation were verified on an Apple Silicon machine; the Intel build is cross-compiled and has not been run on Intel hardware. Whisper.cpp and automatic learning from edits are Windows-only for now.
What about the ChatGPT and Doubao “no key” providers?
They reuse a session you sign into yourself inside the app and talk to undocumented endpoints. They are off by default, may break at any time and may not comply with those services' terms — the account risk is yours to judge. Prefer the offline engine or your own API key if unsure. See DISCLAIMER.md.
Can I use it without any internet at all?
Yes. After the one-time model download (resumable, SHA-256 verified, three mirrors), everything runs offline. The portable build keeps its data next to the executable.
Something went wrong — where do I report it?
Open an issue on GitHub. Failed recordings are kept locally (max 20 clips / 7 days) so you can retry from History without re-speaking, and the main log lives in the data folder.
Stop typing. Start speaking.
Free, open source, offline by default. Install in a minute and keep your voice to yourself.
v0.17.2 · MIT · Source on GitHub