Voice

Two halves, and you can have either without the other.

• The Studio hears you. Press the microphone, say it, and the words land in the message box.
• The agents speak. An agent says a line or two aloud before it writes — or reads you its whole answer.

Both use a speech provider you bring your own key for, the same way models do. Both are set up on one page: Settings → Studio → Speech.

What it needs#

A key for a speech provider. Add one under Settings → Models → Providers, in the Speech providers section: Deepgram, Fish Audio or ElevenLabs. Saving checks the key by speaking one word with it.

For hearing only, you can also use a chat connection you already have — OpenRouter or Venice — with no separate key.

Talking to the Studio#

The microphone sits beside Send — just after it in the chat box, and before the picture button in the Studio agent and in a plugin's chat. Click it — or press Ctrl+Shift+L from anywhere — and the message box turns into a live wave with a clock.

Speak, then stop in any of these ways:

• Click the microphone again (Stop: hear what I said).
• Just stop talking. After a few seconds of quiet the clip ends on its own — five by default, and yours to set from one to ten.
• Press Esc or the × to drop the clip. Nothing is sent and nothing is billed.

While it works out what you said it reads Hearing… — click again to cancel. Then the words appear in the box, ready to edit and send. Turn on Send what I say and they go straight out instead, as if you'd pressed Enter; if the agent is still busy, your message waits its turn.

It's push-to-talk: there's no wake word and it never listens on its own. The words come all at once when you're done, not as live captions. A clip runs two minutes at most, and if you say nothing at all for twenty seconds it gives up with Heard nothing.

The first time, your system may ask for permission to use the microphone. If it's refused, the box says The microphone was refused. Allow it for Semantix Studio and try again. If it hears only silence, the box says so and points at your system's sound settings. That's usually the wrong input device, not Semantix.

The settings, under Hearing: The Studio hears you (on by default — the button appears as soon as a provider can hear), the Hearing provider and its Model, the Language you'll speak (this window's language by default, or detect; some providers always detect and don't show the row), Quiet that ends it, and Try it to test the whole thing.

Hearing the agents#

Turn on Agents speak. Then choose what The voice reads:

• A short spoken version — the agent says one to three sentences aloud first, then writes its full answer. What you hear and what you read aren't meant to match word for word.
• The full answer — the written answer is read to you as it arrives, minus code blocks, tables, links and other things that don't survive being read aloud.

Pick the Speech provider and a Voice, and press Try the voice to hear it.

Any finished answer has a small speaker beside it — Read this answer aloud — to hear it again, or for the first time.

Muting: the Voice menu in the header has Mute the voices. Muted, nothing is sent to the provider and nothing is billed; the agent's spoken line still appears in the chat, marked with a small speaker, so you don't miss what it said.

Speaking, like typing, interrupts: press the microphone and any voice that's playing stops.

A voice per agent#

Each of your agents can sound like itself. Under Project Manager → Your Agents, an agent's Voice card has Speaks with — The Studio's voice, or any voice from a provider you have a key for — and Hear it to try it. Where the voice supports them, Speaking pace and Expressivity fine-tune it.

An agent with a voice of its own only ever uses that voice. If its provider loses its key, the agent goes quiet and says why. It never borrows another agent's voice.

Presence#

When an agent speaks, something moves: its presence, a small animation behind the conversation. It's large on the Studio agent's card before the first message, and a band above the conversation after.

It moves for two reasons only — the voice actually sounding, and a signal: a memory being recalled, or an event arriving — and fades to a faint outline at rest, so you know where it will come back. It has no moods and no "thinking" animation.

Choose it on the Theme tab → Visual, under Look: Sphere (the default), Waveform, Rings, Ribbons, Nebula, or None. Try it plays a sample. Every look runs live underneath, so you can compare them. The presence saves with your theme.

Where it is now#