You talk. It types. Local voice typing for Windows: hold a key, speak, release, and the text appears in the app you are already using.
Speech recognition runs on your machine. There is no account, no subscription and no telemetry. It reaches the internet only when you ask it to: to fetch a speech model, to look for an update if you have allowed that, or for the optional Claude rewrite, which is off until you turn it on and give it a key. Nothing you dictate leaves this machine unless that last one is on.
Hold Right Ctrl (or whichever key you pick), talk, let go. A second or two later the text appears where your cursor is — in your editor, your browser, a chat box, anywhere.
- Two local engines. NVIDIA Parakeet TDT 0.6B v3 for English and 24 other European languages, OpenAI
Whisper
large-v3-turbofor the rest. TawkType picks per language, or you can force one. - Four modes, switched with a key. Literal keeps what the recogniser produced, Clean prose tidies it, Chat stays lower-case and unspaced, Technical protects acronyms and your own identifiers. All four work with no model involved — picking one never sends anything anywhere.
- It fits the text to where it lands. Before typing, it reads the few words either side of your cursor: no doubled spaces, a separating space where one is needed, and no capital letter dropped into the middle of a sentence. It only ever acts on what it can actually see.
- It knows where it can type. Before typing it checks whether the focused element actually accepts text, and whether the target window is running as administrator — synthetic keystrokes to an elevated window are discarded by Windows without any error. If it can't type, the text goes to your clipboard and the bar says Copied — ready to paste.
- Your clipboard is borrowed, not taken. Long or multi-line text is pasted rather than typed, and that needs your clipboard for a moment. Everything on it is copied aside first — the image, the file you copied in Explorer, the formatting on copied text, not just the plain words — and put back afterwards. One thing does not survive: a file that is not really a file, such as a mail attachment or something inside a zip, because Windows hands those over as a stream that cannot be copied aside. Ordinary files are fine. If you copy something yourself while a dictation is landing, yours wins and TawkType leaves it alone.
- Nothing is lost. Every dictation is kept locally, so one that went into the wrong window is a click away. Turn the log off and it means it — nothing keeps a second copy of your words.
- History that corrects things. Search what you have said, edit a past dictation, run cleanup over it again, delete one entry, or turn a mistake into a permanent correction — TawkType works out which words actually changed and offers to remember just those.
- A dictation box when delivery fails. If the text could not be typed, it waits in an editable scratchpad you can copy from, correct, or send back to the window it was aimed at. It never opens itself over what you were doing — the bar tells you, and you open it when you want it.
- Hold, or toggle. Hold the key for a sentence; set an optional second key that starts and stops with a press each, for long passages or when holding is awkward. Esc abandons either — nothing is typed, nothing is recorded.
- Your own words. Names TawkType should spell your way, corrections for what it keeps mishearing, and saved text you insert by saying "insert" and a trigger. Each one is added through a dialog with real fields, so there is no punctuation to remember and a multi-line snippet is typed as several lines rather than an escape code. Any of the three lists can still be opened as plain text to paste a list in or sort one, and the whole vocabulary imports and exports as a file of its own. All of it works on this machine, with or without a Claude key.
- Sounds and a safety net. Optional start, finish and failure sounds, off by default and using your own Windows scheme. A recording limit — five minutes by default — finishes a runaway dictation rather than throwing it away, so a key left under a book costs you nothing.
- You can see where your words go. The bar carries a Local or Cloud badge while you work, and its tooltip spells out the whole state: whether recognition is local (it always is), whether the Claude rewrite will run, and whether anything is being written to disk. It reports what will actually happen rather than what the settings imply — switch the rewrite on without an API key and it still says Local, because nothing will be sent.
- Optional AI cleanup. With a Claude API key, dictations are rewritten before typing: spoken corrections applied ("no, make that Tuesday"), lists formatted. Your vocabulary is re-applied afterwards, so a rewrite can never undo a correction you wrote down. The bar says Rewriting with Claude while it happens, because that is the one step that leaves your machine.
Download TawkTypeApp-win-Setup.exe from the latest release and run it. It installs per-user and needs no administrator rights. It can update itself from that same release feed, but it does not look for updates unless you let it: Check for updates automatically on Settings → General is off by default, and Check now is there for when you want to look.
The installer carries an
Appsuffix the application does not. Velopack names it after the package id, and that id has to differ from%LOCALAPPDATA%\TawkType— the folder your settings and models live in — because the installer wipes its own folder on every update.
Windows will warn you the first time. TawkType is not code-signed yet, so SmartScreen shows "Windows protected your PC". Choose More info → Run anyway. That warning is about the absence of a paid certificate, not about anything found in the file.
The installer is around 37 MB. Speech models are not included and are not fetched automatically — nothing that large is downloaded without you asking for it. The first run walks you through choosing one, and through everything else, ending with you dictating a sentence.
Or run from source.
The first time TawkType starts it opens a short setup, and the last thing it asks you to do is dictate. Seven steps:
- Welcome — what it does, and what it does with what you say.
- Microphone — pick your input device and watch the level move as you talk.
- Model — choose a language; TawkType says which engine that needs and what it costs to download, then fetches it and loads it.
- Key — hold the combination you want to dictate with. It warns you about the obvious clashes: the shortcuts nearly every application has, the Windows key, Caps Lock, and a plain key that would still type while you held it.
- Try it — hold your key, say a sentence, and watch it appear in a box TawkType owns.
- Privacy — whether to keep a history, and whether to allow the Claude rewrite.
- Done — what is actually true, and an offer to start with Windows.
Each answer is saved as you give it, so there is nothing to confirm at the end and nothing to lose if you close the window. Nothing is taken on trust either: you cannot move past the microphone step until the meter has moved, past the model step until the model has loaded rather than merely downloaded, or past the practice step until a dictation has actually produced words. "Ready" means it worked.
Skip setup is there if you cannot finish today, and the --setup flag opens the walkthrough
again whenever you want it.
- Windows 10 or 11, 64-bit
- A microphone
| Minimum | Comfortable | |
|---|---|---|
| CPU | 2 cores | 4+ cores |
| RAM | 4 GB | 8 GB |
| Free disk | ~1 GB | ~3 GB if you keep both models |
| GPU | none needed | 2 GB VRAM, for Whisper |
A GPU is optional. Whisper uses Vulkan when a current graphics driver is present and falls back to the CPU. Parakeet is CPU-only and fast enough there. No CUDA Toolkit, no Python, no Rust.
The default engine, and the one those figures are about.
- 640 MB on disk, and about 850 MB of memory while it is loaded — that is the whole application, not just the model.
- CPU-only, using half your logical cores, never fewer than two and never more than eight. On an eight-core machine it takes four, so there is headroom for whatever you are dictating into.
- Measured on an i7-11700F: 3.1 s to load at startup, then 931 ms to transcribe 13 seconds of speech.
Those timings are one desktop. A two-core laptop has fewer threads, slower ones, and less memory bandwidth, so expect several seconds to load and something nearer real time to transcribe — an estimate, not a measurement.
The alternative, for the ~75 languages Parakeet does not cover. It behaves quite differently.
- 1.5 GB on disk, and with a GPU the weights live in video memory, not system RAM: measured at ~1.5 GB of VRAM, while the application itself sits at about 330 MB resident.
- ~2.9 s to load on Vulkan — close to Parakeet's, despite the model being more than twice the size, because the GPU is doing the work.
- Needs roughly 2 GB of free VRAM. On a machine with no usable GPU it falls back to the CPU, where those weights move into system RAM instead — expect around 2 GB, and slower. That fallback is reasoned rather than measured.
So the two engines cost different things: Parakeet spends system RAM and CPU, Whisper spends VRAM and very little else. If you have a graphics card, Whisper is cheaper on the machine than it looks.
Running from source needs the .NET 8 SDK; the installer needs only the .NET 8 Desktop Runtime, and offers to fetch it.
git clone https://github.com/mburns67priv/TawkType.git
cd TawkType
dotnet run --project src/TawkType.AppTawkType lives in the system tray and on the taskbar, and opens the first run the
same way an installed copy does. Parakeet is about 640 MB, Whisper large-v3-turbo about 1.5 GB, and
they go into %LOCALAPPDATA%\TawkType\models.
After that, hold Right Ctrl and talk.
It sits on screen, dimmed, showing your hotkey; the moment you speak it comes back to full strength with a live level meter and a timer. Its toolbar has Settings, History, Copy last dictation and the dictation box — which lights up when a dictation is waiting in it. A Local or Cloud badge and the active mode's name sit on the right: a mode you cannot see is a mode you will be surprised by, and the same goes for where your words are going. Drag the bar anywhere; it remembers where, per monitor.
Prefer it solid rather than faded between dictations? Settings → Appearance.
It never takes focus. The window is WS_EX_NOACTIVATE, so Windows delivers your clicks but never
activates it. Press a button or drag it and your caret stays exactly where it was.
Eight pages, light or dark or following Windows:
| Page | What's there |
|---|---|
| General | Overview, your dictation stats, where the files live, deleting your data, and what happens to it if you uninstall |
| Transcription | Engine, language, microphone, and managing downloaded models |
| Activation | Push-to-talk key, optional toggle key, tap threshold, how text gets inserted, sounds, recording limit |
| Modes | The four behaviours, the key that cycles them, and what each one does |
| Appearance | Theme, and where the status bar sits |
| Vocabulary | Spellings, replacements and snippets — with import and export |
| AI cleanup | The optional Claude rewrite |
| History | The log's settings; the dictations themselves live in the History window |
Pick a hotkey by clicking the box and holding the keys you want, rather than typing their names. A number that a box cannot use says so underneath, and Save waits until it is fixed — nothing is dropped silently or clamped to something you did not choose.
The built-in cleaner strips "um" and fixes spacing. It cannot tell that "the deadline is Monday, no wait, make that Tuesday" should come out as "The deadline is Tuesday." That needs a model.
Turn it on in Settings and paste an Anthropic API key. The key is
encrypted with DPAPI under your Windows account in apikey.dat — never in settings.json.
ANTHROPIC_API_KEY works too.
When it's on, the transcript — not the audio — is sent to the Anthropic API. If the call is slow (2 s by default), fails, or you have no key, the plain cleaned-up text is typed instead, so a dead network degrades dictation rather than breaking it.
The prompt is explicit that the transcript is speech to be typed, never an instruction. Dictate "write me a poem about the sea" and you get that sentence, not a poem.
Nothing, unless you ask for it. Speech recognition, the phrase book, the modes, the caret fitting and the history are all local, always — there is no account, no subscription and no telemetry.
TawkType connects to the internet in exactly three situations, and you choose every one of them:
| When | What it sends | |
|---|---|---|
| Downloading a speech model | You pick one in the first run or in Settings → Transcription. Nothing large is ever fetched without being asked for | Nothing about you |
| Checking for updates | Off by default. Either you switch on Check for updates automatically in Settings → General, or you press Check now | Nothing about you — it reads a public list of releases |
| The Claude rewrite | Off until you turn it on and supply a key | The transcript, your vocabulary and your custom instructions |
The rewrite is the only one of the three that sends anything you said. With it on, each dictation sends the transcript of what you said, your vocabulary list, and any custom instructions you have written to the Anthropic API. It does not send the audio, the contents of your screen, the text around your cursor, or your history.
Settings → General shows the current state in plain sentences and carries both switches — the history and the rewrite, independently. The same summary is on the status bar's badge as you work, and the bar says Rewriting with Claude during the one step that leaves your machine.
Windows' own cloud clipboard is an exception TawkType closes for you. If you have clipboard sync switched on, anything that reaches your clipboard is uploaded to your Microsoft account and pushed to your other devices — and dictations reach your clipboard, both when a long one is pasted and when one that could not be typed is copied for you. TawkType marks everything it writes as do not upload, so that does not happen. The transient text used for a paste is kept out of the local clipboard history (Win+V) as well; text copied for you to paste yourself stays in it, because you are going to need it.
The badge describes what will happen rather than what is ticked. Three things have to be true before anything is sent: the rewrite is on, the active mode permits it (Literal never does), and a key is stored. If any one of them is false the badge says Local and tells you which.
Everything is under %LOCALAPPDATA%\TawkType\ — deliberately a different folder from the
application itself, which the installer puts in %LOCALAPPDATA%\TawkTypeApp\. The installer wipes its
own folder on every update, so keeping your data out of it is what stops an update erasing your
settings and models:
| File | What |
|---|---|
settings.json |
All settings. Plain text |
apikey.dat |
Your Anthropic key, DPAPI-encrypted for your Windows account |
history.jsonl |
Every dictation. Plain text — turn it off in Settings → History if that's not for you, and the file is deleted rather than merely hidden |
models\ |
Downloaded speech models |
logs\tawktype.log |
Debug log: two files, a little over 10 MB at most, then the older is overwritten. Records how long and how many characters, never the words themselves. Turn it off, or clear it now, in Settings → General — the history switch does not cover it |
Uninstall TawkType the ordinary way: Settings → Apps → Installed apps, find TawkType, and choose Uninstall. It is a per-user install, so no administrator prompt appears. The Start Menu entry, the tray icon and the application folder all go.
Your data does not. By default an uninstall leaves %LOCALAPPDATA%\TawkType\ exactly as it is —
your models, your history, your vocabulary, your settings and your encrypted key. That is deliberate:
a speech model is up to 1.5 GB to download again, and your vocabulary is your own work. Reinstall
later and TawkType picks up where you left off, with no download and nothing to set up a second time.
If you would rather it all went, you have two ways to do it, and which one you want depends on whether you are leaving or tidying.
Before uninstalling, open Settings → General and tick Delete all of this if I uninstall TawkType. Then uninstall normally. The data folder is removed as part of it, with nothing further to confirm.
It has to be decided in advance rather than during the uninstall, and that is not an oversight: the installer's hooks are not permitted to put a question on screen, and are stopped if they take longer than half a minute. Asking beforehand is the only way the question can honestly be asked at all — so if you skip this step and uninstall, your data stays behind.
Settings → General → Delete my data… removes the same things immediately, while TawkType keeps running. Use it when you want a clean slate, when you are handing the machine on, or when you meant to tick the box above and are about to uninstall.
Before it deletes anything it shows you the total size and names what is about to go — how many models, how many dictations in the history, whether a key is stored, and the exact folder. Nothing happens until you say yes, and it cannot be undone.
Afterwards TawkType still works. It simply starts again from nothing: the next dictation needs a model, so it has to be downloaded a second time, and the first run offers to walk you through it again.
If TawkType is already gone and you want to be sure, delete %LOCALAPPDATA%\TawkType in File
Explorer — paste that path into the address bar.
Nothing outside that folder belongs to TawkType except one registry value: TawkType under
HKCU\Software\Microsoft\Windows\CurrentVersion\Run, written only if you turned on Start TawkType
when I sign in to Windows. TawkType removes it during the uninstall, whatever you chose about your
data — an entry pointing at a deleted application is nobody's idea of a feature. If you want to
check, it is the row named TawkType in Task Manager → Startup apps.
Measured on an i7-11700F with an RTX 4060 Ti, same 13-second clip, best of 5:
| Engine | Runs on | Load | Transcribe |
|---|---|---|---|
| Parakeet TDT 0.6B v3 (int8) | CPU, 8 threads | 3.9 s | 931 ms |
| Whisper large-v3-turbo | GPU (Vulkan) | 2.4 s | 413 ms |
Both transcripts were identical and correctly punctuated. Typical end-to-end: 3.2 s of speech typed in 215 ms.
dotnet test # 416 unit tests, ~2 s
dotnet run --project tools/TawkType.Bench -- speech.wav Both 5
dotnet run --project tools/TawkType.Clean -- "um the deadline is monday no wait tuesday"
dotnet run --project tools/TawkType.Focus -- 15 # what the focus probe sees
dotnet run --project tools/TawkType.Brand # regenerate the icon and logosLaunch flags: --settings, --history, --dictation-box, --setup, --overlay-demo.
The version comes from the last release tag — git describe — so a local build reports the release it
descends from rather than a number someone forgot to bump. -p:Version still wins where it matters.
To build an installer locally:
./build/pack.ps1 -Version 0.2.9Releases are cut by tagging: git tag v0.2.9 && git push origin v0.2.9 runs
.github/workflows/release.yml, which tests, packs and publishes the GitHub Release that installed
copies update from.
docs/ARCHITECTURE.md explains the design and docs/HANDOFF.md is the working notes, including the
gotchas that cost the most time. docs/REVIEW-2026-09-11.md and docs/FEATURE-RESEARCH-2026-09-11.md
are point-in-time assessments, each carrying a note of what has been addressed since.
TawkType leans on other people's work:
- NVIDIA Parakeet TDT 0.6B v3 — © NVIDIA Corporation, used unmodified under CC BY 4.0, via the ONNX int8 export by csukuangfj
- OpenAI Whisper — MIT
- sherpa-onnx (Apache-2.0) and Whisper.net (MIT) run them
- Wispr Flow is the product this is modelled on
Full list, with licences: THIRD-PARTY-NOTICES.md.
Apache-2.0. Open source: use it, modify it, redistribute it, build something else out of it, commercially or otherwise. Keep the copyright and licence notices, say which files you changed, and the patent grant in section 3 protects everyone including you.
The source is public partly so the privacy claims above can be checked rather than believed, and partly so anyone who wants to take this further can.
Contributions are welcome. Apache-2.0 section 5 means anything you deliberately submit for inclusion arrives under the same licence, so there is no CLA to sign. Open an issue first for anything large, so nobody writes something that was never going to be merged.
Third-party components keep their own licences — see THIRD-PARTY-NOTICES.md and NOTICE.
tawktype.com · built by Michael Burns.