AI Dictation

How Trueears captures your voice and delivers formatted text into any app

Trueears captures your voice via a global shortcut, transcribes it through Groq Whisper, optionally formats the result with an LLM, and pastes the final text directly into whatever app you’re using — all in under three seconds.

The Dictation Pipeline

Every recording follows the same path from hotkey press to pasted text:

1. Hotkey detected

   You press Ctrl+Shift+K. The Rust backend’s shortcuts.rs module intercepts this as a system-wide global shortcut — no need for Trueears to be focused.

2. Window detection

   Before the overlay appears, window.rs calls the OS to identify the foreground window (Win32 GetForegroundWindow on Windows, xdotool on Linux). The app name, window title, executable path, and cursor position are captured and sent to the frontend via a Tauri IPC event.

3. Overlay shown

   The transparent, always-on-top overlay window appears near your cursor, displaying a recording indicator. The overlay is click-through when inactive so it never interrupts your workflow.

4. Audio recording

   The frontend starts capturing audio using the browser’s MediaRecorder API inside the Tauri WebView. Audio stays local — it never passes through the Rust backend.

5. Stop triggered

   You stop recording by pressing Ctrl+Shift+K again, releasing the key (Push-to-Talk), or pressing Escape to cancel.

6. Groq Whisper transcription

   The audio blob is sent directly from the frontend to the Groq Whisper API (whisper-large-v3-turbo by default). The raw transcription text is returned, typically within a second.

7. LLM post-processing (optional)

   If LLM formatting is enabled, dictationController.ts matches the active window against your App Profiles, selects the appropriate system prompt, and sends the raw transcription to Groq Chat for formatting. The LLM is instructed to format — never to respond conversationally.

8. Auto-paste

   The frontend calls the transcription_complete Tauri command. automation.rs writes the text to the clipboard using arboard and simulates Ctrl+V with enigo, pasting directly into the original app.

Recording Modes

Configure your preferred mode in Settings > Preferences.

Mode	Behavior	Best For
Auto (default)	Quick tap = Toggle; hold = Push-to-Talk	Maximum flexibility
Toggle	Press once to start, press again to stop	Long dictation sessions
Push-to-Talk	Hold to record, release to stop	Short commands and quick notes

Auto mode is the most versatile: a brief tap behaves like Toggle for longer passages, while holding the key activates Push-to-Talk for quick one-liners.

See Recording Modes for detailed configuration instructions.

The Overlay

The overlay is a transparent, always-on-top window that spans all monitors. Key design properties:

• Cursor-positioned — appears near wherever your cursor is, not in a fixed corner
• Click-through when idle — when not recording, all mouse events pass straight through
• No focus stealing — the overlay never takes focus away from the app you’re dictating into
• Visual indicator — shows an animated recording state so you always know when the mic is live

On Wayland (Linux), the overlay uses a smaller centered window with set_focusable(false) to replicate this behavior within portal constraints.

Transcription Model

Trueears uses whisper-large-v3-turbo via the Groq API by default. This model provides the best balance of speed and accuracy for real-time dictation.

To change the model:

1. Press Ctrl+Shift+S to open Settings
2. Go to the Transcription tab
3. Select a different Whisper model from the dropdown

Groq provides a free tier with generous limits — most users will never exceed it for normal dictation use.

LLM Post-Processing

The optional LLM formatting step sends your raw transcription through Groq Chat before pasting. The LLM receives a system prompt that instructs it to:

• Clean up filler words and disfluencies
• Apply formatting appropriate for the active app (e.g., bullet points in Notion, professional tone in Outlook)
• Never respond conversationally — it outputs only the formatted version of what you said

If the LLM returns something that looks like a refusal ("I cannot...", "As an AI...", etc.), Trueears automatically falls back to the raw transcription.

To enable LLM post-processing:

1. Open Settings (Ctrl+Shift+S)
2. Go to the LLM Post-Processing tab
3. Toggle the feature on and enter your API key
4. Select a model (default: openai/gpt-oss-120b)

Context-aware formatting is driven by App Profiles — see App Profiles to learn how Trueears adapts its output per application.

Performance Targets

Metric	Target
Hotkey press to recording start	< 100ms
Transcription displayed after speech ends	< 3s

Actual transcription time depends on audio length and Groq API response time. The whisper-large-v3-turbo model is optimized for low latency.

Keyboard Shortcuts

Action	Windows / Linux	macOS
Start / stop recording	Ctrl+Shift+K	Cmd+Shift+K
Open Settings	Ctrl+Shift+S	Cmd+Shift+S
Cancel recording	Escape	Escape

Recording Modes

Configure Auto, Toggle, and Push-to-Talk in detail

App Profiles

Control how the LLM formats text per application