Using voice mode

Once you’re in voice mode, the panel changes to a screen built for talking rather than reading. What it looks like depends on the engine you picked.

On GPT-Live

Connecting. The moment you switch voice on, the panel becomes one full surface with an orb in the middle and “Connecting…” underneath. A second or two later the voice says hello. It uses your name if it knows it, and varies the greeting with the time of day rather than saying the same line every time.

The orb. The orb turns with whoever’s speaking, and the colour that leads tells you which of you it is. When nobody’s talking it turns slowly. When your agent is off doing something for you, it slows right down and a short line underneath says what’s happening.

Just talk. There’s nothing to press. Speak, and it answers. If it’s mid-sentence and you’ve heard enough, talk over it; it stops and listens. If it’s answering and you want to add something, add it.

Three controls, at the bottom of the panel:

• Transcript flips to the written conversation so you can read along or scroll back. Flip back whenever you like.
• Mute silences your microphone until you press it again, for when someone walks in.
• End hangs up and returns the tab to text. The conversation carries on from there as a normal chat.

When it hands something to your agent. The voice is the one you’re talking to. It can chat, answer from what’s been said, and keep you company. It can’t open your files or run a task. The moment you ask for something that needs real work, it hands that to your agent in the same tab, tells you it’s on it, and comes back with what your agent found. In the transcript you’ll see the request marked ↗ Handed to Claude, then what your agent did, then what the voice said about it. The next page explains this in more detail.

On ElevenLabs

The blob. In the middle is a blob that changes colour depending on what’s happening: listening, thinking, speaking. A glance tells you whether it’s waiting for you, working, or about to reply.

Just talk. Start speaking. It listens, and when you’ve finished a thought, it goes to work and replies out loud. Then it’s listening again. You don’t press anything between turns.

Flip to the transcript any time to read along, and back again. Nothing’s lost either way; it’s the same conversation shown two ways.

Leaving voice

On GPT-Live, press End. On ElevenLabs, switch the tab back to text from the microphone in the status bar. Either way the conversation carries on as a normal chat, and because voice is per-tab, your other conversations are untouched.

On your phone

Voice works on your phone too, and it follows whichever engine your Mac is set to. On GPT-Live your Mac sets up the call for you, so you don’t need a key on the phone for calls; just keep Obsidian open on the Mac with mobile access on. On ElevenLabs, hold or tap the mic to talk and Hyo speaks back.