Privacy & data

What Ditto does and doesn't send to the internet, what it stores on your machine, and what other apps can see.

The short version

• Audio is captured, processed, and discarded locally. It never leaves your machine.
• Transcriptions are generated by whisper.cpp and optionally formatted by Qwen3 through llama.cpp. Both run on your computer.
• Ditto talks to the internet only to download Whisper or Qwen models from HuggingFace, on demand, when you click Download.
• No telemetry, no analytics, no crash reports phone home.

Audio

When you record:

1. The renderer captures audio with getUserMedia and MediaRecorder — the same APIs a website uses for a “record voice” feature.
2. The audio is decoded, resampled to 16 kHz mono, and converted to 16-bit PCM in the renderer process.
3. The PCM is sent to the main process over IPC as an ArrayBuffer.
4. Main writes it to a temporary .wav file in %TEMP%.
5. whisper-cli.exe reads the WAV, transcribes, prints the text to stdout.
6. The temp .wav is deleted in the finally block of the IPC handler, immediately after transcription.

The audio never crosses a network boundary. The only time it touches disk is the temp WAV, which is removed right after use. If the app crashes mid-transcription, leftover WAVs are cleaned up by Windows’ temp folder maintenance.

Transcriptions

Once Whisper produces text, Ditto:

1. Optionally passes it to the local Qwen formatter for punctuation, paragraphs, lists, and filler removal.
2. Saves the final text to local history when history is enabled.
3. Sets your clipboard to the final text.
4. Optionally simulates Ctrl+V into the active app.
5. Optionally restores your previous clipboard content.

The formatter server listens only on a temporary authenticated 127.0.0.1 port. Text is not sent to a remote API.

Models

Ditto downloads Whisper and Qwen models from HuggingFace’s CDN on demand:

• The download happens only when you click Download in Settings → Models or in the welcome window.
• Whisper models go to %APPDATA%\ditto\models\; the Qwen formatter goes to %APPDATA%\ditto\formatter-models\.
• Once downloaded, models are never re-fetched. Ditto doesn’t check for updates or phone home about them.

The models are served by HuggingFace’s standard CDN. HuggingFace may keep its own download access logs; that is separate from Ditto and contains no audio or transcription text.

Settings and other local data

Ditto stores everything related to your configuration in:

%APPDATA%\ditto\config.json

This is plain text JSON, managed by electron-store. You can open it in any editor to inspect or back up. The schema matches the DittoSettings type in the code.

What’s in there:

• Your shortcut, theme, mic, language preferences
• Whether smart formatting and local history are enabled
• Cumulative usage counters (transcribe count, total audio seconds processed, number of auto-pastes)
• Last pill drag position, if you have “remember last position” on

What’s not:

• Any audio
• Any account or identifier — Ditto has no concept of accounts

Transcription text is stored separately in ditto-history.json only while local history is enabled. Entries contain final text, creation time, selected model and language, and whether formatting was applied.

Telemetry, analytics, crash reports

There is none. Ditto does not include:

• An analytics SDK
• A crash reporting service (Sentry, Bugsnag, etc.)
• An auto-update channel that pings home (auto-update is planned, but you’ll be able to opt out, and it’ll only contact GitHub Releases when checking for updates)
• Any network call other than model downloads from HuggingFace

If a future release adds any of those, it will be opt-in and disclosed in the release notes.

Microphone permission

Windows treats microphone access as a sensitive permission. The first time Ditto records, Windows pops up a permission prompt. Ditto needs that to work — without it, no audio reaches the app.

You can review or revoke the permission at any time:

• Windows Settings → Privacy & security → Microphone
• Find Ditto in the list, toggle access on/off

Ditto only uses the mic during an active recording. There’s no background listening, no wake-word detection, no continuous capture. The getUserMedia stream is requested when you press the shortcut and stopped the moment you press it again.

Clipboard

Ditto reads the clipboard only to snapshot it for the “Keep previous clipboard” feature, and only when both auto-paste and that option are on.

Ditto writes to the clipboard when transcription finishes — your text goes there before the simulated Ctrl+V (or as the only effect, if auto-paste is off).

That’s the entire interaction Ditto has with the clipboard. It doesn’t read the clipboard outside of recordings. It doesn’t store clipboard history.