Whisper voice input

Whisper lets you talk to the agent instead of typing. You speak, and each phrase becomes a message in the chat. Recognition runs on your computer with whisper.cpp, so your voice is never sent anywhere to be transcribed, not to AgenC and not to a provider. Only the text goes to the chat's model, which may be a cloud provider.

Use it

1. Open a chat, or stay on the welcome screen.
2. Click the microphone in the composer ("Start Whisper"), or the Whisper button at the top right of the chat ("Turn on Whisper. Send after a pause.").
3. Speak. A moving wave shows the app is listening.
4. Pause. After a short silence, 1.2 seconds by default, the phrase is transcribed on your computer and sent as a message.
5. Click the stop button in the composer ("Stop Whisper") to stop.

What you already typed in the composer is left alone. Sent phrases go through the same queue as typed messages, so you can keep talking while the agent works.

On the welcome screen, turning Whisper on first prepares an empty session with your chosen folder, model and permissions, without sending anything, and starts listening there.

The first time, your operating system asks for access to the microphone. On macOS the request says "AgenC uses the microphone when you record a voice prompt or explicitly enable local Whisper dictation."

When it stops listening

The app never listens in the background. The microphone turns off, and any unfinished audio is discarded, when you:

• stop it, or switch to another chat;
• leave the chat, open Settings or hide or minimize the window;
• reconnect Core, or change a Whisper setting that affects recording.

After a reload or a restart, Whisper starts off. A phrase still being transcribed is never delivered to another chat.

The speech model

Whisper needs a speech model on your computer. The first run of the app downloads the Base model, about 141 MB, and setup waits for it. The Small model, about 465 MB, is more accurate; download it in Settings > Whisper > Speech model. Core checks each model's size and checksum before it uses it, and keeps it in its data folder. The speech model never replaces your chat model.

Settings

Settings > Whisper holds every option. The top of the page says "Audio stays on this computer. Transcribed text is sent to your chat model."

Waveform

• Wave style: Aurora, Tide, Ember, Prism or Signal. The style shows inside the composer's stop button while you speak.
• Intensity: Subtle, Balanced or Expressive, how much the wave moves.

Whisper mode

• Send after a pause: "Automatic". Each phrase is sent after the silence delay.
• Silence delay: 0.8, 1.2 (default) or 2 seconds of quiet before a phrase is sent.
• Microphone sensitivity: Low, Normal (default) or High. "High responds to quieter speech; low ignores more background sound."

Transcription

• Local engine: whisper.cpp, and whether it is "Ready on this computer." Refresh checks again.
• Speech model: Base or Small, with a Download button for a model you do not have yet.
• Spoken language: Auto detect, or one of English, Spanish, French, German, Italian, Portuguese, Dutch, Polish, Russian, Ukrainian, Chinese, Japanese, Korean, Arabic, Hindi and Turkish.
• Output language: Keep original language, or Translate to English.
• Vocabulary hints: up to 500 characters of names, acronyms and specialist terms that help recognition. "Names and terms only, never secrets." They are passed to the engine as a command argument while it runs, so other programs on this computer could see them.
• Acceleration: Auto lets whisper.cpp use the hardware it finds; CPU only does not use the graphics chip.

An older Core supports fewer of these options. The page then says "Update Core for more languages, translation, vocabulary hints, and acceleration controls." When your choices need more than that Core offers, it also shows Use compatible settings.

Limits

• One phrase is transcribed at a time, with up to three waiting. Long speech is split into pieces of about 20 seconds.
• If the chat refuses a finished phrase, the text is kept as an "Unsent voice message", in the chat and in Settings > Whisper, with Retry send, Copy text and Discard. The app never resends it on its own.
• Whisper is not available from a phone or another browser.
• To attach a recording instead, see Writing messages.