Configure voice calls

Voice Calls lets customers talk to a public agent on supported web surfaces. Configure these settings with care. Calls use credits, and the system can record them.

Enable voice

1. Open an agent. Then select Channels → Voice Calls.
2. In Configuration, choose a provider, language, and maximum call duration.
3. Save your changes. Enable voice. Then place a test call from the public experience.

Supported pinned languages are English, Spanish, French, German, Portuguese, and Japanese. You can set the maximum call duration from 1 minute to 60 minutes. An organization can have up to 25 active calls at the same time.

Choose a provider

The Provider control holds three tabs. An agent uses one tab at a time. Each row shows the rate in credits per minute and the equivalent price per minute.

Realtime uses one speech-to-speech engine for the whole call. Eight engines are available, from 1 to 12 credits a minute. One engine carries a Recommended badge.

The Realtime list responds to the language you pinned. AgentStack holds measured accuracy data for some engines and some languages only. When an engine was measured as the best for the pinned language, that engine moves to the top of the list and carries a Most accurate badge. An engine that was measured as weak for that language shows a warning instead. An engine with no measured data for that language carries no badge, which is not a statement about its accuracy.

Combinations uses a chain of three engines: speech-to-text, then an LLM, then text-to-speech. Select a named chain. The name is what the call bills under, and it is what the call history shows. Eight named chains are available:

• Zero-Cost Speech — 1.5 credits a minute. Its text-to-speech stage bills nothing, so only transcription and reasoning cost credits.
• Budget Mix — 2 credits a minute.
• LiveKit House — 2 credits a minute. Budget Mix with a larger reasoning stage.
• Multilingual — 2 credits a minute. For callers who may not speak English.
• Warm Voice — 2 credits a minute. An expressive speech stage.
• Full Grok — 2 credits a minute. xAI at every stage.
• Balanced Mix — 3 credits a minute.
• Deep Reasoning — 6 credits a minute. The strongest reasoning stage.

Build your own selects one engine for each stage: 5 speech-to-text engines, 8 LLMs, and 4 text-to-speech engines. The rate of the chain is the sum of the three stages, from 1.5 to 6 credits a minute. A chain can cost a fraction of a credit each minute.

If your chain is the same as a named chain, the agent saves that name. The call then bills under the name. Any other chain bills as a custom combination.

Warning: a change of provider changes the price of every later call. Check the rate on the row before you save.

Configure recording and privacy

Open Recording & Privacy. Decide whether to record calls.

If you enable recording, configure the disclosure that customers see. Then make sure that it matches the consent requirements that apply to you.

Review a call

Use Call history to review completed calls. Audio and transcripts are protected resources. Share them only with authorized workspace members.

You can also review a call from the agent's Activity page. Set the channel filter to Voice. Open a session to view the call length, the credits it used, and a player for its recording. Call turns are not saved as messages, so a call-only session shows an empty transcript. See Activity.

Troubleshoot

• Call button missing: Make sure that the agent is public. Make sure that voice is enabled for that surface.
• Call ends immediately: Make sure that credits are available. Make sure that the provider configuration is correct.
• Microphone unavailable: Grant microphone permission in the browser. Then retry the call from a secure HTTPS page.
• Another call is already active: Before you start a new call, end or replace the existing session.

Voice calls use the same workspace balance as other usage. For details, see Billing and credits.