Whisper Dictation for Windows: Setup Guide
Press a hotkey, speak, and press it again: your words are typed into whatever window you're working in. Speech is turned into text by Whisper running on your own PC, so nothing is sent to the internet.
The latest version of this guide is always at https://tastysnak.gitlab.io/whisper-dictation/.
This guide works on any reasonably modern Windows PC, with or without a graphics card. The one choice that depends on your PC is which AI models to use, and the guide shows you how to work that out. It doesn't give a single answer, because the right model on a gaming PC is the wrong one on a laptop.
How it works
- Tap F9 to start recording, then tap it again to stop. Or hold F9 while you speak and let go when you're done. A small label at the bottom of the screen shows when it's recording and when your text has been pasted. You can change F9 to another key.
- Whisper turns your speech into text.
- Optional cleanup: a small local AI model, run by Ollama, tidies the text. It removes "um"s and repeated words, applies self-corrections ("Tuesday, no, Thursday" becomes "Thursday") and fixes punctuation.
- The text is pasted into the window you were in when you started speaking, and your clipboard is put back afterwards.
- Every dictation is saved to a history log, so nothing is lost if it lands in the wrong place.
There's no separate Whisper program to install. Whisper is built into ffmpeg, a free audio/video tool, and the dictation script drives ffmpeg.
What you need
| Requirement | Notes |
|---|---|
| Windows 10 or 11 | |
| PowerShell 7 | Free; installed in step 2 |
| ffmpeg 8.0 or newer, full build | Free; installed in step 3. Older versions and "essentials" builds don't include Whisper |
| A microphone | Any mic Windows can see: USB, headset, webcam or audio interface |
| A Whisper model file | Free download, 75 MB to 1.6 GB. You choose it in step 5 |
| Ollama (optional) | Free; needed only for the cleanup step (step 6) |
| A graphics card (optional) | Makes everything much faster. AMD, NVIDIA and Intel all work, through Vulkan. Without one, everything runs on the processor (CPU) and you use smaller models |
Setup
1. Find out what your PC has
The models you can use depend mostly on your graphics card and how much video memory (VRAM) it has. AI models must fit in that memory to run fast.
- Press Ctrl+Shift+Esc to open Task Manager, and click Performance.
- Click each GPU entry and note:
- the card's name, top right;
- Dedicated GPU memory, its total. That's your VRAM.
Look at the totals:
- Dedicated GPU memory of 2 GB or more: you have a graphics card that will help.
- Only "Shared GPU memory", or a GPU with well under 1 GB dedicated: the GPU is built into the processor. Plan as if you have no graphics card.
- No GPU listed: you have no graphics card.
Memory is shared with everything else. Games, video editing and other AI tools use the same VRAM. If you run those at the same time as dictation, leave them room. Pick models that fit in, say, half your VRAM rather than all of it.
You'll get exact measurements in step 7. This step is only so you know which models to download first.
2. Install PowerShell 7
Open Terminal or Command Prompt and run:
winget install Microsoft.PowerShellWindows has an older "Windows PowerShell 5.1" built in. The script needs version 7, which is called
pwsh.
3. Install ffmpeg
winget install Gyan.FFmpegClose and reopen your terminal, then check that ffmpeg has Whisper:
ffmpeg -hide_banner -filters | findstr whisperYou should see a line containing whisper A->A Transcribe audio using whisper.cpp. If nothing appears, your ffmpeg is too old or isn't a full build. Run winget upgrade Gyan.FFmpeg.
4. Create the dictation folder and download the script
Open PowerShell 7: search for pwsh in the Start menu. Then create a folder called whisper-models in your user folder and download the dictation script, dictate.ps1, into it:
New-Item -ItemType Directory -Force "$env:USERPROFILE\whisper-models"curl.exe -L -o "$env:USERPROFILE\whisper-models\dictate.ps1" https://tastysnak.gitlab.io/whisper-dictation/dictate.ps1You can also download dictate.ps1 in your browser and move it into the folder. It's a plain text file; open it in Notepad to see exactly what it does before you run it.
The folder can live anywhere. The script looks for models and its settings file in its own folder.
5. Choose and download a Whisper model
Whisper models come in sizes. Bigger ones make fewer mistakes, but they're slower and need more memory. The aim is the most accurate model that still feels instant on your PC.
| Model file | Download | Memory while transcribing | Typical fit |
|---|---|---|---|
ggml-tiny.en.bin |
75 MB | ~0.3 GB | Very old or low-power PCs. Noticeably more mistakes |
ggml-base.en.bin |
142 MB | ~0.4 GB | No graphics card, or laptops |
ggml-small.en.bin |
466 MB | ~0.9 GB | A fast CPU, or a graphics card with 2–4 GB of VRAM |
ggml-large-v3-turbo.bin |
1.6 GB | ~1.9 GB | A graphics card with 4 GB of VRAM or more. Excellent accuracy and still fast |
Skip the others. medium is slower than large-v3-turbo and less accurate. large-v3 is only slightly more accurate than turbo, but needs about twice the memory and is several times slower.
Models ending in .en understand English only, and are slightly better at it. If you'll dictate in another language, use large-v3-turbo, or ggml-base.bin / ggml-small.bin without the .en.
What to download first:
- Graphics card with 4 GB or more:
large-v3-turboandbase.en. The benchmark in step 7 compares the two, and the small one is a fallback. - Smaller graphics card, or none:
base.enandsmall.en.
Download straight into the folder. Run one or more of these, one for each model you want:
large-v3-turbo (1.6 GB):
curl.exe -L -o "$env:USERPROFILE\whisper-models\ggml-large-v3-turbo.bin" https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v3-turbo.binbase.en (142 MB):
curl.exe -L -o "$env:USERPROFILE\whisper-models\ggml-base.en.bin" https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.en.binsmall.en (466 MB):
curl.exe -L -o "$env:USERPROFILE\whisper-models\ggml-small.en.bin" https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-small.en.binFor any other model, change the file name in both places. All the models are at https://huggingface.co/ggerganov/whisper.cpp/tree/main.
With several models in the folder, the script uses the most capable one it finds, in the order large-v3-turbo, large, medium, small, base, tiny, unless you choose one in the settings. Step 7 tells you whether that's really the right one for your PC.
6. (Optional) Set up AI cleanup
Cleanup needs Ollama, a free tool for running AI models on your own PC, plus one chat model.
winget install Ollama.OllamaHow big a cleanup model should be. Our tests on dictation text found two things:
- Models of about 4 billion parameters ("4B") are the sweet spot. They need about 2.5–4 GB of memory and take a fraction of a second per sentence on a graphics card.
- Smaller isn't safer. 1B–2B models sometimes answer what you dictated instead of typing it. "What's the capital of France" came out as "Paris", for example. They also miss more filler words.
- Bigger isn't better. A 20B model needed more than three times the memory, was about ten times slower, and was no more accurate. It also couldn't tidy long passages in time.
| Model | Memory | Test results |
|---|---|---|
qwen3.5:4b |
~3.8 GB | Best overall. Never answered instead of typing; removed every filler; handled 5-minute passages |
gemma3:4b |
~3.0 GB | About as good. Better at self-corrections, slightly worse at fillers |
bonsai-2:4b |
~2.4 GB | The smallest that was still safe. A little weaker on long sentences |
Whisper and cleanup need memory at the same time. Add them up and compare with your VRAM from step 1:
| Your graphics card | Suggested start |
|---|---|
| 12 GB or more | large-v3-turbo + qwen3.5:4b (~6 GB total). Plenty of room left |
| 8 GB | large-v3-turbo + qwen3.5:4b, or gemma3:4b / bonsai-2:4b if you also run games or other AI tools |
| 6 GB | small.en + bonsai-2:4b, or large-v3-turbo with cleanup off |
| 4 GB or less, or none | Start with cleanup off. Dictation works well without it. Try bonsai-2:4b only if the benchmark shows it's quick enough |
Install the one you want to try:
ollama pull qwen3.5:4bIf Ollama isn't installed or running, dictation still works and simply skips cleanup.
7. Measure: run the benchmark
This step replaces guesswork with measurements on your PC. The benchmark:
- tests every Whisper model in the folder and every chat model in Ollama;
- tells you which ones to use;
- changes nothing.
Run:
pwsh -ExecutionPolicy Bypass -File "$env:USERPROFILE\whisper-models\dictate.ps1" -BenchmarkIt shows your graphics card, then a short paragraph. Press Enter, read the paragraph aloud the way you'd normally dictate, and press Enter again when you've finished. It takes about a minute, more with several cleanup models.
Here is a real run on a PC with a 20 GB graphics card. To make it repeatable, the paragraph was read by Windows' built-in computer voice:
Graphics: AMD Radeon RX 7900 XT, 20 GB of video memory (VRAM)Speech: 15.7 secondsModel Size Time Word errors Fast enough?ggml-base.en.bin 141 MB 0.5 s 0.0% yesggml-large-v3-turbo.bin 1549 MB 1.2 s 2.4% yesGPU check: the GPU is being used (ggml-base.en.bin is 3.5x faster on the GPU than on the CPU).Model VRAM On GPU Per sentence Passed Failedqwen3.5:4b 3.6 GB 100% 0.25 s 5/5gpt-oss:20b 12 GB 100% 1.80 s 5/5=== Suggestion ===Whisper model: ggml-base.en.binCleanup model: qwen3.5:4bThe computer voice is perfectly clear, so even the small model got every word right, and the benchmark suggested it. With a real voice, accents, a fast pace or a noisy room, the bigger models usually pull ahead. That's why you should do the reading yourself, in the room where you'll dictate.
Reading the Whisper table:
- Time is how long that model takes to turn your recording into text. It includes loading the model, because dictation does that every time.
- Word errors is the share of words it got wrong compared with the paragraph you read. Under about 5% is very good. Over 15% means you'll be correcting a lot.
- Fast enough? means it finished within a quarter of the time you spent speaking. Slower than that, and you'll notice the wait.
- GPU check says whether Whisper really runs on your graphics card. If it says the GPU is not used, update your graphics driver. Until then, stay with the small models.
The suggestion is the most accurate model that's fast enough. If a smaller model is nearly as accurate, within 2 points, it picks the smaller one: it's quicker and leaves more memory free. One short reading can be a little lucky or unlucky, so if two models are close, run the benchmark again and see whether the result holds.
Reading the cleanup table:
- VRAM is the memory the model actually used.
- On GPU below 100% means the model didn't fit in video memory, and part of it runs on the much slower CPU. Pick a smaller model, or close other programs.
- Per sentence is the typical time it adds to each dictation. Under 1 second feels fine.
- Passed shows how many of five quick checks it got right: removing fillers, applying a self-correction, leaving a question unanswered, ignoring an instruction hidden in the text, and keeping a request that's meant for someone else.
The suggestion is the smallest model that passed all five within 1 second. It names nothing if no model managed that. In that case, try another 4B model, or turn cleanup off.
If it suggests something bigger than you'd like, because you need the memory for other things, choose a smaller model that still says "yes" and passed 5/5. The tables give you what you need to trade speed and memory against accuracy yourself.
Run it again whenever you download another model, change your graphics card or update its driver.
A non-English speaker can't use the built-in paragraph. Record yourself speaking for about 15 seconds in any app, then run
-Benchmark -File "path\to\recording.wav". You'll get the times and the GPU check, but no word-error score.
8. Put your choices in the settings
The first run created a settings file called dictate.json in the script's folder. Open it in Notepad and set the two models the benchmark suggested. The benchmark prints the exact lines to use:
"Model": "C:\\Users\\YourName\\whisper-models\\ggml-large-v3-turbo.bin",..."Cleanup": { "Enabled": true, "Model": "qwen3.5:4b", ... }In the settings file, backslashes in paths are written twice.
Name the cleanup model rather than leaving it on "auto". "auto" picks the smallest chat model installed in Ollama, which might be a tiny one that doesn't do the job well. To turn cleanup off, set "Enabled": false.
9. Check your microphone and test
The script uses your Windows default microphone, which you set in Settings → System → Sound → Input. To see what it detects:
pwsh -ExecutionPolicy Bypass -File "$env:USERPROFILE\whisper-models\dictate.ps1" -ListMicsThen test. Stay quiet for the first 3 seconds while it measures background noise. When it says TEST: recording 2 seconds..., say a few words:
pwsh -ExecutionPolicy Bypass -File "$env:USERPROFILE\whisper-models\dictate.ps1" -TestYour words appear after ->, with the time each step took.
10. Create shortcuts
Run this once in PowerShell. It puts a Dictation shortcut on your desktop, and another in your Startup folder so dictation starts by itself when you log in:
$script = "$env:USERPROFILE\whisper-models\dictate.ps1"$pwsh = (Get-Command pwsh).Sourceforeach ($folder in [Environment]::GetFolderPath('Desktop'), [Environment]::GetFolderPath('Startup')) { $lnk = (New-Object -ComObject WScript.Shell).CreateShortcut("$folder\Dictation.lnk") $lnk.TargetPath = "$env:SystemRoot\System32\conhost.exe" $lnk.Arguments = "--headless `"$pwsh`" -NoProfile -ExecutionPolicy Bypass -File `"$script`" -Background" $lnk.WorkingDirectory = Split-Path $script $lnk.IconLocation = "$pwsh,0" $lnk.Save()}The shortcuts start dictation without a window: it lives in the system tray by the clock (see Daily use). If you don't want it to start at login, press Win+R, type shell:startup, press Enter and delete the shortcut there.
Ollama also starts at login. If it isn't ready yet, dictation waits up to a minute for it before starting without cleanup.
Daily use
- Start Dictation, or let it start at login. A small label appears at the bottom of the screen.
- Stay quiet for about 3 seconds while it says
Calibrating. It measures your room's background noise each time it starts, so it adapts to fans, open windows and other noise. - When it says Dictation ready, these keys work:
| Key | Action |
|---|---|
| F9 (tap) | Start recording; tap again to stop and paste |
| F9 (hold) | Push-to-talk: records while held, pastes when you let go |
| F10 | Quit |
| F11 | Turn start/stop beeps on or off |
| Shift+F11 | Turn AI cleanup on or off |
Click where you want the text before you start speaking. The text goes to the window that was active when recording began, even if you click somewhere else while talking.
Knowing what it's doing
The microphone icon by the clock shows that dictation is running:
| Icon | Meaning |
|---|---|
| Grey ring | Starting up |
| Blue-grey | Ready: tap or hold your dictation key |
| Red | Recording |
| Amber | Turning your speech into text |
| Orange with ! | Something's wrong: right-click it and choose Show log |
Windows may hide new icons under the ^ arrow. To keep this one visible, drag it from there onto the taskbar.
The label at the bottom of the screen appears only while you dictate:
- ● Recording 0:04 · tap F9 to stop, or release F9 to stop if you're holding the key;
- Transcribing..., then Tidying up... if AI cleanup is on;
- ✓ Pasted, or Nothing heard, and then it fades away.
It never takes the keyboard focus and clicks go straight through it, so it can't change where your text goes.
Right-click the icon for:
| Menu item | What it does |
|---|---|
| AI cleanup | Turn cleanup on or off (the same as Shift+F11) |
| Beeps | Turn start/stop beeps on or off (the same as F11) |
| Change dictation key... | Pick a different key instead of F9. Press the key you want, with Ctrl, Alt or Shift if you like, and click Use this key. It works straight away and is saved for next time |
| Open history folder | Your saved transcripts and recordings |
| Show log | Everything dictation has printed since it started, including the raw transcript and how long each step took |
| Quit dictation | The same as F10 |
On its own, a letter, number or Space would stop you typing it, so the key picker only accepts those with Ctrl or Alt held. Function keys, Pause, Scroll Lock and Insert work alone.
History
Every dictation is added to a daily log in the history folder next to the script, for example history\2026-11-03.txt:
[14:02:11] I think we should meet on Thursday to go over the budget. raw: Um, so I think we should, uh, meet on Tuesday, no wait, Thursday, to go over the budget. audio: 2026-11-03_140211.wavThe recording is kept too, so you can transcribe it again later, for example with a bigger model:
pwsh -ExecutionPolicy Bypass -File "$env:USERPROFILE\whisper-models\dictate.ps1" -File "$env:USERPROFILE\whisper-models\history\2026-11-03_140211.wav"Recordings older than 30 days are deleted automatically. The text logs are kept.
Settings (dictate.json)
The file is created on first run in the script's folder. Edit it in any text editor, then restart dictation.
{ "Mic": "auto", "Model": "auto", "ModelDir": "", "Language": "en", "SilenceDb": "auto", "HoldMs": 400, "PasteTarget": "start", "Beeps": false, "Hotkeys": { "Toggle": "F9", "Quit": "F10", "Beeps": "F11", "Cleanup": "Shift+F11" }, "Cleanup": { "Enabled": true, "Model": "auto", "Url": "http://localhost:11434", "TimeoutSec": 15, "KeepAlive": "30m" }, "History": { "Enabled": true, "Dir": "", "KeepAudio": true, "KeepAudioDays": 30 }}| Setting | What it does |
|---|---|
Mic |
"auto" uses the Windows default microphone. Or an exact name from -ListMics |
Model |
"auto" uses the most capable Whisper model in ModelDir. Or the full path to a model file, which is better once you've run the benchmark |
ModelDir |
Folder to search for models. "" means the script's own folder |
Language |
Language you'll speak, e.g. "en", "de", "es", or "auto" to detect it. Use a model without .en for languages other than English |
SilenceDb |
"auto" measures background noise at every start. Or a fixed number such as -40 to skip calibration (see Troubleshooting) |
HoldMs |
Holding the toggle key longer than this many milliseconds means push-to-talk; a shorter press is a tap |
PasteTarget |
"start" pastes into the window active when recording started. "current" pastes into whatever window is active when it finishes |
Beeps |
true plays a beep when recording starts and stops |
Hotkeys |
Any key, optionally with modifiers: "F9", "Pause", "Scroll", "Ctrl+Alt+D", "Ctrl+Shift+Space". "" disables the Beeps or Cleanup key. The dictation key (Toggle) is easiest to change from the tray menu: Change dictation key... |
Cleanup.Enabled |
false turns AI cleanup off permanently |
Cleanup.Model |
The Ollama model name, e.g. "qwen3.5:4b". "auto" picks the smallest installed chat model, which may not be a good one |
Cleanup.Url |
Where Ollama is running. Change it to use Ollama on another computer, for example one with a bigger graphics card |
Cleanup.TimeoutSec |
If cleanup takes longer than this, the raw text is used |
Cleanup.KeepAlive |
How long Ollama keeps the model in memory between dictations. Shorter frees memory sooner. The first dictation after that takes a few seconds longer while the model loads again |
History.Enabled |
false turns the history log off |
History.Dir |
Where history is saved. "" means a history folder next to the script |
History.KeepAudio |
false saves only the text, not the recordings |
History.KeepAudioDays |
Recordings older than this are deleted. 0 keeps them forever |
Command-line options
These override the settings file for one run:
| Option | Effect |
|---|---|
-Benchmark |
Time and score your models and suggest the best ones (step 7). Changes nothing |
-Test |
Record 2 seconds, print the transcript, exit |
-File <audio> |
Transcribe an audio file, print it, exit (no mic needed). With -Benchmark, benchmark that file |
-ListMics |
Show the default mic and every mic the script can use |
-Mic "name" |
Use this microphone |
-Model <path> |
Use this Whisper model file |
-Language <code> |
Use this language |
-NoCleanup |
Skip AI cleanup this run |
-Background |
For starting without a console window (the shortcuts use it): errors appear in a message box, and the output goes to dictate.log |
Started from a terminal without -Background, dictation shows the same tray icon and label, and also prints everything in the terminal. This is handy for testing.
Troubleshooting
Transcription is slow
Run -Benchmark. If the GPU check says the GPU isn't used, update your graphics driver. Otherwise, choose a smaller model that the benchmark marks "yes" and put its path in "Model".
"GPU check: the GPU does NOT seem to be used" Whisper is running on the CPU:
- Update the driver. Install the latest graphics driver from AMD, NVIDIA or Intel, restart, and run the benchmark again.
- Built-in GPU only. A GPU that is part of the processor often doesn't help much. Use
base.enorsmall.en. - Built-in GPU and a graphics card. Many PCs have both, and the benchmark lists both. The GPU check shows whether the speed-up is there. If the check passes but transcription is still slow, compare your times with the example in step 7, which comes from a high-end card. If yours are many times slower than you'd expect for your card, Whisper may be using the built-in GPU. Update both drivers and run the benchmark again.
Cleanup is slow, or "On GPU" is under 100% The cleanup model doesn't fit in your graphics card's memory next to everything else. Try each of these:
- A smaller 4B model:
gemma3:4borbonsai-2:4b. - Close games or other AI programs.
- Turn cleanup off with Shift+F11.
It always says "(nothing heard)" Your speech is quieter than the silence gate. Check:
- Calibration: were you quiet during the 3-second calibration? Restart dictation and stay silent.
- Mic volume: is it too low? Raise it in Sound settings → Input.
- Fixed gate: still failing? Set a fixed value,
"SilenceDb": -45. Lower numbers (more negative) let quieter sound through.
It types "Thank you." or other phrases you didn't say
Whisper sometimes invents phrases when it hears only noise. The script drops the common ones automatically. If it keeps happening, your background noise may be getting past the gate. Set a fixed SilenceDb a little higher (less negative), e.g. -35.
AI cleanup changed what I meant
- Check what you said: the history log keeps the raw transcript under every cleaned one, and so does Show log in the tray menu.
- Turn it off: press Shift+F11 to turn cleanup off for a while, or set
"Cleanup": { "Enabled": false }. - Built-in safety check: if the cleaned text looks very different from what you said, for example because the model answered your question instead of tidying it, the script pastes the raw text instead.
- If it happens often, run the benchmark and pick a model that passes all five checks.
"AI cleanup is off" notification, or "Cleanup: off (Ollama not reachable)" in the log Ollama isn't running. Start it from the Start menu (it normally runs in the system tray), then restart dictation.
"Cleanup: off (Ollama model '…' not installed)"
The model named in Cleanup.Model isn't installed. Run ollama pull <name>, or change the name in the settings.
Is it running? Look for the microphone icon by the clock. Check under the ^ arrow too. No icon means it isn't running: start it from the Dictation shortcut.
"Dictation is already running" Only one copy can run at a time. Use the one that's already in the tray, or right-click its icon, choose Quit dictation, and then start it again.
"Could not register F9 (already used by another app?)"
Another program has taken that key; games, recording software and some keyboards do this. Close the other program, or pick a different key: set Hotkeys.Toggle in dictate.json, then start dictation again. Once it's running, you can also change the key from the tray menu, which tells you straight away if a key is taken.
Hold vs tap gets confused
If short presses start push-to-talk, or long presses don't, adjust HoldMs: a higher number makes push-to-talk need a longer hold.
Wrong microphone
Set the right one as default in Windows Sound settings, or put its exact name (from -ListMics) in "Mic".
Text doesn't appear in one particular program Windows blocks normal programs from typing into programs running as administrator. Run dictation as administrator too, or use that program normally.
"ffmpeg not found" or "no whisper filter" Repeat step 3, then open a new terminal. Programs only see newly installed tools after they're restarted.
"running scripts is disabled on this system"
Always start the script with -ExecutionPolicy Bypass, as in the commands and shortcuts above.
Privacy
After the one-time model downloads, everything runs on your PC, including the AI cleanup, which uses Ollama locally. Nothing is uploaded anywhere. The only exception is if you point Cleanup.Url at another computer yourself.
The history folder keeps your transcripts and recordings. Recordings are deleted after 30 days and the text logs are kept. Delete the folder whenever you like. To keep nothing, set "History": { "Enabled": false }.