Voice input
Dictate into the composer instead of typing. Ekorbia records from your microphone and transcribes your speech on your machine — the audio is never uploaded.
Voice input is macOS-only for now.
Dictating
- Click the microphone button in the composer (between the prompt button and the model picker).
- Speak. The button turns red and shows a timer while it’s recording.
- It stops on its own a moment after you finish speaking — or click again to stop immediately. Your words are transcribed and inserted at the cursor, ready to edit or send. (Prefer to always click to stop? Turn off “Stop automatically when I pause” in Settings → Voice.)
Press Esc while recording to cancel without transcribing.
Voice input is also available in the quick-query overlay — the same mic button sits in the overlay’s input row, so you can dictate a quick question without switching to the main window. The overlay stays open while you’re recording.
You can also trigger overlay dictation from anywhere with the global voice hotkey (default ⌘⇧V on macOS; rebind under Settings → General → Voice input) — it opens the overlay and starts listening immediately, so you can dictate without switching to Ekorbia first.
By default Ekorbia sends the message as soon as a dictation finishes — speak and it goes, with no Enter or Send click (especially handy in the overlay: speak a question, get an answer). Turn off “Send automatically after dictation” in Settings → Voice if you’d rather review before sending; pressing Esc while recording always cancels without sending.
The first time you record, macOS asks for permission to use the microphone — click Allow. If you dismiss it or change your mind later, enable Ekorbia under System Settings → Privacy & Security → Microphone.
Choosing a speech model
The first time you use voice input, Ekorbia offers to download a small speech model. Three English models are available:
| Model | Size | Best for |
|---|---|---|
| base.en | ~142 MB | The recommended default — fast and accurate |
| tiny.en | ~75 MB | The fastest option / lowest memory |
| small.en | ~466 MB | The most accurate, a little slower |
For languages other than English, download a multilingual model instead —
base, small, or large-v3-turbo (most accurate, ~1.6 GB). These cover 99
languages and can also translate to English.
Models download once and then run entirely offline. On Apple Silicon transcription is GPU-accelerated and typically takes well under a second for a short dictation.
Manage your speech models any time under Settings → Voice — download another, switch which one is the default, or remove ones you don’t use.
Other languages
By default Ekorbia transcribes English. To dictate in another language, download a multilingual model (see above), then open Settings → Voice:
- Language — choose Auto-detect or a specific language. Auto-detect works well for a sentence or more of clear speech.
- Translate to English — when on, your speech is transcribed and translated into English in one step.
Both settings apply only to multilingual models; the English-only (*.en)
models always produce English.
Notes
- Everything runs locally. The only network use is the one-time model download from Hugging Face; after that, voice input works with the network off.
- Dictation works in private chats too — it’s just another way to enter text.
- Speech models are separate from your chat (Ollama) models: the model picker in the composer chooses which LLM answers; the voice model only turns your speech into text.
- If you see “No audio captured,” the microphone permission likely isn’t granted yet — check System Settings → Privacy & Security → Microphone.