Voice conversations

Talk with your agent#

In an empty chat input, choose the voice waveform button to start a live conversation. When you type or attach a file, that position becomes Send. The separate microphone button dictates an editable text draft instead.

Allow microphone access when your browser asks. On first use, review the credit rate and choose Start with credits. You can ask about your CRM, deals, tasks, or other apps the selected agent can access. Results appear in the same chat, and a pending approval waits above the composer. The agent asks follow-up questions aloud. When an action needs approval, it describes the action and asks you to say approve or reject; an unclear or qualified response does not run the action.

Your words and the spoken reply appear progressively as messages in the chat. The compact card above the input shows your agent’s avatar, which reacts to your microphone as you speak and animates while the agent talks, plus a mute control. It respects reduced-motion preferences. The voice button becomes an X to end the conversation. Pricing appears before your first call, keeping the call controls focused on the conversation.

Within each voice reply, the agent's detailed results appear before the spoken transcript, along with its tool activity when Show the agent's steps is on. Files, approvals, errors, and results that arrive after you end voice remain available. The transcript stays in the chat after the call settles.

Control your conversation#

Use the microphone control to mute yourself and End voice to stop the call. Muting does not stop credit usage. Interrupting spoken audio does not cancel an action already running: use the chat’s Stop control when you want to cancel agent work. Ending voice leaves authorized work in chat.

Voice stays connected while you navigate within Opser. On desktop, a working conversation follows you into the agent sidebar; on mobile, compact controls let you mute or end voice while viewing an app. Idle conversations do not open automatically. A call has no set length: it ends by itself after 10 minutes without anyone speaking. Closing the tab, signing out, losing the connection, or running out of credits also ends it. Reconnecting requires a new call; actions are not replayed automatically.

Keyboard shortcuts#

Focus the chat and use Ctrl+D to start or stop dictation (Control on Mac too), or Ctrl+Shift+Space to start or end live voice (Cmd+Shift+Space on Mac). The shortcuts appear in each button’s tooltip and follow the same controls: voice starts when the input is empty, the first-use credit notice still appears, and disabled controls stay disabled. If two chat panels are open, focus the one you want to use. While you dictate, press Enter to send what you said; Ctrl+D or the pencil puts it in the message box to edit first. Dictation already knows your teammates, agents, apps, and company; to teach it other words or fix one it keeps getting wrong, see Teach dictation your words.

Choose your voice#

GPT Live speaks with your selected Opser agent's name and personality. Voice tool work uses GPT-5.6 Luna with low reasoning, with workspace and dynamic app tools available. Question forms and connection cards stay in text chat because they require on-screen interaction. Text chat keeps its configured model and reasoning. If your agent uses a connected Codex subscription, voice tool work uses Luna through that subscription; otherwise it uses the metered OpenAI API. Live audio is always billed separately in credits.

Open the agent’s General settings and pick a voice under Agent voice. Each agent keeps its own voice, and a call uses the selected agent’s voice. Play the preview beside any voice to hear a short sample without spending workspace credits. Changes apply to your next call. Its owner or an admin can also ask the agent to change it.

Understand credits#

The first-use notice shows your workspace’s current credit rate for live voice. Live voice is billed per second, with a 15-second minimum when connecting. Silence, muted time, and waiting for the agent count while the session is open.

Opser reserves a short amount of credit ahead of use. Unused reserved credits are returned when the call is settled. If a connection fails before final usage arrives, Opser retains the confirmed usage it received and marks the final amount as unconfirmed for billing review.

The agent’s model and tools are charged separately at their usual rates. Live voice and dictation use workspace credits even when your agent’s text model uses your connected ChatGPT subscription. Dictation is billed by the recording’s duration, at its own rate.

Open Settings → Usage to see voice and dictation alongside other model usage and inspect the credit ledger for reservations and refunds. See Credits and usage.

Audio and access#

OpenAI processes the audio. Opser retains text transcripts, not audio recordings. Voice uses the selected agent’s existing access and approval policies. Spoken approval resolves only an explicit approve or reject response; a pending approval shows above the composer, and decided ones stay in the chat when Show the agent's steps is on.

If voice cannot connect, check browser microphone permission and your workspace’s credit balance. Use a browser with WebRTC support; voice availability in an embedded mobile browser depends on its microphone support.