Voice orchestrator settings control how the assistant handles spoken conversations — how it listens, how it speaks, and how it behaves during a call.

What the page shows depends on the voice version the assistant runs on:

• Voice 2.0 runs calls on a native speech-to-speech model (GPT Realtime), which listens and speaks in one step. It speaks the caller's language and can switch language partway through a call. The page has three sections: Instructions, Speech detection, and Select voice.
• Voice 1.0 runs calls on the legacy Retell pipeline, which converts speech to text and synthesizes the reply. It understands and speaks English only. The page has two sections: Transcription settings and Select voice.

Where to find it

Open the Admin Console, then go to Orchestrator → Voice.

Which version is the assistant on?

Enabling voice

Until voice is switched on, the page shows a disabled state — Voice is disabled, with a prompt explaining that enabling it lets employees speak and receive spoken responses. Click Enable voice to turn it on. Voice is always enabled on voice 2.0; there is no version to choose.

To turn it off again, open the more menu (⋮) and choose Disable voice orchestrator. A confirmation appears first, because disabling removes voice calling for the whole organization and it has to be switched back on manually — there is no automatic recovery.

Switching to voice 2.0

An assistant on voice 1.0 moves to voice 2.0 from this page:

1. Open the more menu (⋮) in the page header and choose Switch to voice 2.0. It isn't available while you're editing — save or cancel your changes first.
2. Read the confirmation and click Switch now.

The page reloads with the voice 2.0 settings, and the next call runs on voice 2.0.

What happens to existing settings

If the assistant takes phone calls

Test in UAT before the default changes

Switch your UAT assistant, then run the calls your employees actually make — on web, and on phone if you use it:

• Your highest-volume requests, end to end — leave balance, tickets, policy questions, approvals.
• Requests that carry IDs, codes, email addresses and names. Voice 2.0 has no boosted keywords list, so check that the terms your employees say most are understood.
• Every language your employees speak — on web and phone, starting a call in each one, switching partway through a call, and mixed speech such as Hindi with English words. Check the spoken answers and the chat cards that come with them.
• Call control — interrupting the assistant, cancelling a running request, putting the call on hold and resuming.
• Sign-in prompts during a call, if any of your workflows ask the employee to authenticate.
• Your corporate network and VPN. On voice 2.0, web call audio travels over WebRTC directly between the browser and OpenAI, so test from the networks your employees use.
• Noisy places, with the speech detection threshold at its default first.

Review each test call in Call history — status, disconnection reason and the runs it started — then tune the voice model instructions and the speech detection threshold, and test again. Report anything that doesn't behave as expected to your Leena AI implementation contact.

Editing settings

The page opens read-only. Click Edit to make changes, then Save to apply them or Cancel to discard them. Every section is an accordion and is expanded by default.

Voice 2.0 settings

Instructions

Voice model instructions is free-text guidance for the speech-to-speech model — tone, speaking speed, response style, and how to read out inputs such as alphanumeric codes (for example, digit by digit).

It accepts up to 1,000 words. The field shows a running word count and stops accepting text at the limit.

Instructions here shape how the assistant speaks, not what it does. Requests are still answered by the same orchestrator, tools and knowledge as in chat.

Speech detection

Speech detection threshold sets how clearly a sound must be speech before the assistant treats it as the caller talking. It is a slider from 0.30 to 0.95, in steps of 0.05. The default, 0.70, is marked Recommended and suits most callers.

Select voice

Voice 2.0 uses the speech-to-speech model's own voices. Marin and Cedar are marked Recommended, and Marin is the default.

The table shows a preview button to play a sample, the voice name and its gender. Narrow the list with the gender filter or search by name. Voice 2.0 voices aren't tied to an accent or age — every voice speaks whichever language the call is in — so there is no accent filter.

Use Select voice on a row to make it active. The current voice is marked Voice selected. In read-only mode the table shows just the active voice.

Languages

Voice 2.0 is multilingual by default. There is no language list or setting on this page — the assistant follows the caller.

Language handling has no separate setting. Any extra guidance belongs in the Voice model instructions, and should be tested in each language your employees use.

Handled automatically on voice 2.0

These behaviours have no setting:

Voice 1.0 settings (deprecated)

These settings apply only to assistants still on voice 1.0, and don't exist on voice 2.0.

Transcription settings

These tune the quality and behaviour of speech-to-text during a call.

Boosted keywords

Use this for anything the default model is unlikely to know: product codes, internal acronyms, team or system names, employee ID formats. It is the single most effective fix when the assistant consistently mishears a specific term.

You can add up to 100 keywords. Beyond that, the field asks you to remove existing entries before adding more.

Additional instructions

Free text, up to 1,000 words. The field shows a running word count and stops accepting text at the limit.

Small talk frequency

Set in seconds, from 3 to 8, adjusted with the stepper controls rather than typed. The default is 3.

Select voice

Choose the voice the assistant speaks in. Voices span a range of accents, genders, and age styles, so you can match the assistant to your users.

The section header shows how many voices are available. Narrow the list using the gender and accent filters, or search — search matches on accent and age as well as name.

Use Select voice on a row to make it active. The current voice is marked Voice selected. In read-only mode the table shows just the active voice rather than the full list.