Audio Transcription
Kognar turns speech into text for both the microphone button in the chat input and the Voice Assistant. Three transcription paths are available, and you choose how they are used.
Choosing a transcription engine
Open Settings (Ctrl/Cmd+U) → General → Voice & transcription → Transcription engine.
| Option | Behavior |
|---|---|
| Automatic (recommended) | Uses the best available path: macOS native → local Whisper → cloud. No setup needed. |
| Local Whisper only | Always uses local Whisper. Requires the model download below, and falls back to cloud if local transcription fails. |
| Cloud only (Gemini) | Always sends audio to Gemini through your chat connection. No model download needed. |
The setting applies to voice input in the chat and to the Voice Assistant alike.
The three paths
macOS native
On macOS, Kognar can use the system's own speech recognition through a bundled helper. It is fast, needs no download, and keeps audio on the device. Used first in Automatic mode.
Local Whisper (on-device)
Audio is processed entirely on your computer with a Whisper model — nothing goes to the internet.
- Full privacy — audio never leaves your device
- Works offline once the model is downloaded
- Handles longer recordings
- Detects the spoken language automatically; it is not locked to the interface language
Cloud (Gemini)
Audio is sent to Gemini over your existing chat connection.
- No download, works immediately
- Requires an active account and an internet connection
Voice model size
For local Whisper, pick the model in Settings → General → Voice & transcription → Voice model size:
| Size | Trade-off |
|---|---|
| Tiny | Smallest download, fastest, best for simple speech |
| Base | Recommended balance |
| Small | Best accuracy, largest download |
Preparing the model
Click Download / load model to fetch and load it ahead of your first dictation. A progress bar shows the download; otherwise the model loads the first time you use the microphone.
Below the button, Kognar reports the current state:
- Voice model ready (size) — loaded and usable
- Model not loaded in memory yet — dictate once, or use the button
- Loaded model differs from selection — download again or dictate to switch
Voice model settings apply to the desktop app only.
Automatic fallback
Whichever engine you choose, Kognar falls back rather than failing: Automatic walks down the chain, and Local Whisper only falls back to cloud if the local run fails. That way a dictation still produces text.
At a glance
| Path | Privacy | Internet required | Download |
|---|---|---|---|
| macOS native | Stays on device | No | None |
| Local Whisper | Stays on device | Only to download the model | Yes |
| Cloud (Gemini) | Audio sent to the service | Yes | None |