Speak the way you speak. It types what you meant.
Hold a button, talk, let go. Your words are transcribed on your own machine, and a small local model fixes broken grammar and misheard words before they reach the cursor, in any app.
Staging build is green. Shipping after lunch?
Is the login crash fixed yet?
Wherever there’s a cursor
Other dictation types what it heard. This types what you meant.
The speech model writes down exactly what it hears, mistakes and all. A second model on your machine reads the whole sentence and rewrites it into clear English. Flip the switch to compare.
I already tell him many time but he still never fix the bug.
I’ve already told him many times, but he still hasn’t fixed the bug.
I want to deploy the model on the GP you so the inference more fast.
I want to deploy the model on the GPU so that inference is faster.
Write a email to my boss say I am sick today.
Write an email to my boss saying I am sick today.
…you all please don’t late.
Please don’t be late.
Where you mumbled, it knows.
Each clip is transcribed twice in one GPU call, with and without your accent note. Where they disagree, the fixer sees both. Cost: 0.15 s.
B send me the invoice by Friday
The topic settles the word.
Your last two dictations from the past three minutes ride along, so a sound-alike resolves to what you’ve been talking about.
now pie tone → Python
It won’t put words in your mouth.
If a rewrite looks wrong in any of these ways, it’s thrown away and your raw transcript is pasted instead. One- and two-word dictations are never rewritten.
- changes a number
- grows past 1.6×
- shrinks below 0.5×
- keeps <60% of your words
- answers instead of fixing
- adds paragraphs
- translates
- repeats the last dictation
Teach it your repeat mistakes.
cube control => kubectl
A traffic light is the whole UI.
Mouse side button, or record your own.
Left-click the light for your last five copies.
We tested twelve. You get the two best.
Every candidate ran 28 weak-English sentences: grammar, misheard words, context, traps that must be fixed rather than answered, sentences to leave alone, and numbers. Each was scored on the text that would actually be pasted. Switch between the winners from the tray.
Measured 2026-11-16 on an Intel Arc 130V, bf16, greedy decoding, without corrections.txt. The prompt was tuned on Qwen, so the comparison favours it a little. The full table of 12 and the eval script are in the README.
Nothing leaves your machine. It’s still fast.
For a spoken sentence on a Core Ultra 5 laptop with its built-in Arc GPU. Transcription and the English fix take about a second each.
- No upload
- No account
- No telemetry
- HF_HUB_OFFLINE=1
NVIDIA, Apple and Intel Arc GPUs are picked up automatically, and it falls back to the CPU if the GPU fails.
From button to cursor in seven steps.
One dictation is processed at a time, so quick ones queue instead of fighting over the GPU. The log keeps what was heard next to what was pasted.
- Record 16 kHz
Hold the shortcut. On Windows it’s swallowed, so the side button stops paging your browser back.
- Gate < 0.5 s dropped
Silence is trimmed, quiet speech boosted, and clips with no speech (a fan, a knock) are dropped.
- Transcribe twice ~1 s
Qwen3-ASR-1.7B, with and without your accent note, in one batched call.
- Clean up instant
“one two three” becomes 123, “comma” becomes a comma, filler words go. Then corrections.txt.
- Fix the English ~1 s
The local model gets both hearings plus your recent dictations as context.
- Check instant
Guardrails compare the rewrite with the transcript. When in doubt, the transcript wins.
- Paste Ctrl/Cmd+V
Into whatever window has focus. The text stays on the clipboard too.
Set up in five minutes. Then it’s offline for good.
> windows\setup.bat # the speech model, then the English fixer (~8 GB, once) > .venv\Scripts\python.exe download_model.py > .venv\Scripts\python.exe download_model.py --model Qwen/Qwen3-1.7B > windows\run_dictate_background.vbs
$ cd macos && chmod +x *.sh run_dictate.command && ./setup.sh && cd .. # the speech model, then the English fixer (~8 GB, once) $ .venv/bin/python download_model.py $ .venv/bin/python download_model.py --model Qwen/Qwen3-1.7B $ ./macos/run_dictate.sh
$ sudo apt install python3-venv python3-tk libportaudio2 xclip $ cd linux && chmod +x *.sh && ./setup.sh && cd .. # the speech model, then the English fixer (~8 GB, once) $ .venv/bin/python download_model.py $ .venv/bin/python download_model.py --model Qwen/Qwen3-1.7B $ ./linux/run_dictate.sh
dictate.py --doctor.The honest fine print.
Does anything get sent to the internet?
No. The models are downloaded once from Hugging Face. After that the app runs with HF_HUB_OFFLINE=1: no upload, no account, no telemetry. Audio and text stay on your machine.
Is it tested on macOS and Linux?
Not yet on real hardware. It was built and measured on Windows 11 (Core Ultra 5 228V). The macOS and Linux code paths are written and share the same core. On Linux you need an X11 session, because Wayland blocks global shortcuts and tray icons for ordinary apps.
Which languages work?
Speech recognition covers 52 languages and dialects, including strong Chinese and mixed Chinese-English. The rewrite step is for English only, so turn off Fix my English in the tray if you dictate in another language.
What if the fixer changes my meaning?
The guardrails reject most bad rewrites and paste your raw words instead. The log keeps what was heard next to what was pasted, so you can add repeat mistakes to corrections.txt, or switch the fixer off with one click.
Can I use Gemma 4 instead of Qwen?
Yes, from the tray. Gemma 4 E2B scored 28/28 but needs transformers 5, so it runs as a worker in a second environment (.venv-t5). If it fails to load, the app falls back to Qwen3-1.7B and says why in the log.
Why doesn’t it type into some windows?
Windows won’t let a normal app send keystrokes to one running as administrator. Use Restart as Administrator from the tray for those.
Stop typing. Start saying.
Free, local and private. Your voice and your words never leave your machine.