Skip to content
voicefieldguide

How to choose a voice for an AI phone agent

Updated November 6, 2026

The picks

Press play to hear each voice read this guide's sample line.

These picks compare my fifteen voices. I own them, and ElevenLabs may pay me when you use them.

Dexter·First support audition

Even pacing for instructions and confirmation. It gives a sales pitch less excitement than a dedicated ad voice.

0:08

Reel·Friendly short exchanges

A conversational option for brief service turns. Its matter-of-fact delivery won't carry a dramatic brand message.

0:10

Derek·Formal service notices

Firm delivery for short operational notices. A casual shop's greeting may sound too much like a bulletin.

0:08
Add Dexter to My Voices

I’d audition Dexter first for a customer service AI phone agent, with Reel for friendlier short turns and Derek for formal notices. Test the voice through the actual telephone connection, including an interruption and a misheard number. A pleasing library preview cannot tell you how a whole conversation behaves.

These are candidates from my own voices. Their catalog notes support those roles, but your ElevenLabs Agents account must also allow the voice and model combination. ElevenLabs says voices with live moderation enabled cannot be used in ElevenAgents. Check eligibility before building a workflow around any sample here.

What should the voice make easier for a caller?

A caller usually wants a result. They need to hear the question, answer it, and understand what will happen next. I would score a support audition on those moments before judging whether it sounds friendly.

Give the agent one task, such as checking an order status in a test system. Start with a question that has a short answer. A thirty-second greeting full of company history gives the caller nothing useful to do and delays the first chance to discover a mistake.

Then include a correction. Say the wrong digit and repair it halfway through the next sentence. Can the caller tell which number the system has now accepted? The voice needs to make a confirmation sound complete, and the agent’s logic needs to keep the correct value. Those are separate things to review.

Which of my voices fit different support turns?

Dexter for instructions that need a response

Dexter has the even instruction pacing I’d start with for support. He measured about 190 words per minute in my clips, one of my quickest, but phone turns should be sized by the caller’s task. Leave a real response gap after asking for an order number. Dexter’s limitation is promotional energy. A sales team may hear its pitch as too polite.

Reel for a small shop’s conversational service

Reel is casual and direct. It measured about 195 words per minute in my clips, which suits brief exchanges if the lines are short. I would trim a question before asking the voice to rush it. Its flat response to drama is useful to remember when the same brand also wants a big launch announcement.

Derek for a notice the caller should recognize as official

Derek fits a closure notice or operational update. It has firm broadcast pacing, measured near 180 words per minute in my clips. That formality can distance a caller who expects a relaxed neighborhood business. Test names and opening hours before using it for the whole system.

You can add Dexter to My Voices for an account-level audition. My phone systems page covers recorded menus; a conversational agent also needs the turn-taking checks below.

How should I test latency and interruptions?

Measure from the end of the caller’s words to the beginning of the reply. Record several exchanges. Include a short question, a longer explanation and a correction. The pause includes transcription, reasoning, tool calls, networking and speech generation, so changing the voice alone may leave it almost unchanged.

ElevenLabs distinguishes its content-generation model from its real-time variant. A model used to render a polished preview is not a promise of the same response time in a phone agent. Select a model the agent supports and measure the complete call. I would choose reliable turn-taking over a slightly nicer isolated sentence.

Interrupt during an ordinary reply and during a request for confirmation. Listen for audio that continues after the interruption, a repeated acknowledgment, or a clipped first word when the system resumes. Write down what actually happened. “It felt slow” is hard to fix; “the agent repeated the old number after I corrected it” gives you a test to repeat.

What should a first audition script contain?

Use fictional data while evaluating the voice and flow. Here is a spoken opening for a test service line.

Hello. This is the automated support line for the sample store. I can help you check an order or leave a message for the team. Please tell me which you need. If I read a number incorrectly, say stop and give me the correct digits. I’ll confirm them before continuing. You can ask for a person at any time, and I’ll explain the available handoff options.

After it asks a question, answer with a partial number, then correct yourself. Try a speakerphone in an ordinary room as well as headphones. Don’t feed personal customer details into a test prompt just to make it feel realistic.

The script promises a handoff explanation, so implement that behavior. If staff are unavailable, the agent should explain the actual callback or message route. A reassuring voice cannot repair a promise the system does not keep.

Does an outbound call need a separate review?

Yes. The FCC’s February 2024 ruling treats AI-generated voices as artificial voices under the TCPA. Consent requirements and exemptions depend on the call. Have counsel approve your outbound use, audience and records before placing calls. An inbound support audition does not establish permission to dial a list of customers.

Keep voice selection out of the consent decision. Neither a natural delivery nor a recorded disclosure supplies missing authorization. If the outbound workflow hasn’t been approved, finish the voice evaluation with internal test calls.

When is the agent ready for a limited launch?

I would require successful corrections, an understandable handoff and audible names and numbers across the phone connection. Save the failed calls too. They are better regression tests than a perfect greeting.

Let the first production release handle a narrow task with an observable outcome. Review whether it completed that task without making the caller repeat themselves. Expand the task list after the call evidence earns it.

Producing this for a client or an organization? See Voices for IVR menus, on-hold messages and a professional voicemail greeting.

Related guides

Other jobs where I use the same voices.

Nothing playing

Press play on any voice

↑ ↓ moveEnter openShift Enter playEsc