Choosing the right voice for your brand

Rene, Kabir Goel 
Choosing the right voice for your brand

Picking a voice for your brand’s voice agent is not the same as picking a font. A font sits on a page and waits to be read. A voice talks to your customers, handles their problems, and represents your company in real time — often when the customer is frustrated, in a hurry, or calling from a noisy street.

Most teams pick a voice by listening to a few demos and going with whichever “sounds professional.” That approach works until the voice mispronounces the company name on a live call, or sounds cheerful while telling a customer their claim was denied. This guide covers how to choose a voice that actually fits your brand, your use case, and the conditions your callers are in.

Start with the use case, not the voice catalog

The most common mistake is browsing a voice library before defining what the voice needs to do. A voice that works for a meditation app will fail on a collections call. A voice that sounds authoritative in a boardroom demo may sound robotic over an 8kHz phone line.

Before listening to a single voice, answer these questions:

  • What is the caller trying to accomplish? A caller checking an order status wants efficiency. A caller disputing a charge wants empathy. The voice should match the emotional register of the interaction, not the brand’s overall marketing tone.
  • Where is the caller? Phone calls from cars, offices, and kitchens have background noise. The voice needs to be intelligible at 8kHz over PSTN, not just in a studio recording.
  • What happens when things go wrong? The voice needs to handle apologies, corrections, and escalations without sounding tone-deaf. A voice that can only sound upbeat is a liability when the conversation turns serious.

The three characteristics that matter

1. Intelligibility

Can the caller understand what the voice is saying? This sounds obvious, but intelligibility breaks down in specific, predictable places:

  • Numbers and codes. “Your confirmation code is 8-4-2-1-J” needs to be unambiguous. Some voices blur digits together or read “$1,200” as “twelve hundred dollars” when the caller expects “one thousand two hundred.”
  • Names and domain terms. If your product is called “Sonic” and the voice says “Sah-nik,” that’s a problem. Test with your actual product names, customer names, and industry vocabulary.
  • Accented speech. If your caller base includes non-native English speakers or regional accents, the voice needs to be intelligible to them — and your speech-to-text model needs to understand them. These are related but separate problems.

Sonic handles conventional written forms correctly on its own: prices, dates, phone numbers, and addresses. For product names or terms the model mispronounces, a pronunciation dictionary fixes it once instead of requiring prompt engineering in every call.

2. Emotional register

Every voice has a default emotional register — a baseline tone that colors everything it says. Some voices sound warm and reassuring. Others sound clipped and efficient. Neither is wrong; the question is whether the register matches the interaction.

A few examples from our voice library:

Use caseRegisterWhy
Customer supportCalm, measuredThe caller is often frustrated. A calm voice de-escalates.
Outbound salesWarm, energeticThe caller didn’t ask for this call. Energy keeps them on the line.
Collections / complianceClear, firmLegal disclosures need authority. Warmth undermines the message.
Healthcare schedulingSteady, reassuringThe caller may be anxious. The voice should reduce stress, not add to it.

Cartesia’s Sonic model goes further than most TTS systems: it adjusts its emotional delivery based on the text it reads. If the LLM generates an empathetic response, Sonic sounds empathetic. If the text is a firm denial, Sonic sounds firm. This means the voice design is coupled with the LLM’s generation — the voice follows the content, not a static configuration.

3. Latency

A voice that takes two seconds to start speaking after the caller finishes their turn feels broken, no matter how good it sounds. In conversation, people expect a response within about 200 milliseconds. Above one second, callers start repeating themselves or talking over the agent.

Latency is a property of the TTS model, not the voice itself. But it constrains which voices you can use. A voice that requires heavy processing to sound natural may not work in a real-time voice agent, even if it sounds perfect in a demo.

Stock voices vs. cloned voices

Most brands should start with a stock voice. Stock voices are pre-built, tested, and available immediately. Cartesia’s voice library includes voices designed for specific use cases — customer support, sales, scheduling, and more.

Voice cloning makes sense in two situations:

  1. Brand continuity. If your company already has a recognized voice — a spokesperson, a podcast host, a CEO who does earnings calls — cloning that voice maintains consistency across channels.
  2. Differentiation. If every competitor in your space uses the same three stock voices, a cloned voice sets you apart.

Cartesia offers two cloning options: instant voice cloning from 5 seconds of audio, and professional voice cloning from 30+ minutes for higher fidelity. Instant cloning is fast and works well for prototyping. Professional cloning captures more nuance and is better for production use.

A framework for choosing

Here’s the process we recommend:

Step 1: Define the use case. Write down the top three tasks the voice agent handles. For each, note the emotional register, the vocabulary it needs to handle, and the consequences of a miscommunication.

Step 2: Shortlist voices. Pick three to five voices from the library that match the register you need. Don’t overthink this step — you’ll narrow down in testing.

Step 3: Test with real scripts. Write out the ten most common things the agent will say, including the hard ones: apologies, denials, number-heavy confirmations, and product names. Generate audio for each voice and listen to them back to back.

Step 4: Test in context. If the agent will be on the phone, listen to the audio over a phone speaker, not studio headphones. If callers will be in noisy environments, add background noise. The voice that sounds best in a quiet room may not be the voice that sounds best in a car.

Step 5: Pilot with real callers. Run the voice in production for a week. Listen to recordings. Pay attention to the moments where the caller asks the agent to repeat itself — those are intelligibility failures. Pay attention to the moments where the caller’s tone changes — those are register mismatches.

Common mistakes

  • Picking a voice that sounds like the founder. The founder’s taste is not the customer’s experience. Test with real callers.
  • Using the same voice for every use case. A company with both a support line and a sales line may need different voices for each. The brand is the company, not the voice.
  • Ignoring the STT side. The voice the agent speaks with is only half the equation. If the speech-to-text model can’t understand the caller’s accent, the best voice in the world won’t save the conversation. Ink, our streaming STT model, handles accented speech and noisy environments — but test it with your actual caller base.
  • Optimizing for the demo. A voice that impresses in a sales demo may not hold up over a ten-minute support call. Long-form listening reveals artifacts and tics that short clips hide.

Hear your brand's voice in seconds.

Try Sonic free

FAQs

What is a brand voice in text-to-speech?

A brand voice in TTS is the specific synthetic voice a company uses across its voice agents, IVR systems, and audio content. It includes the voice's accent, pace, emotional register, and how it handles domain-specific vocabulary like product names and industry terms. The right voice makes the brand recognizable and trustworthy over the phone; the wrong one creates a disconnect between how a company presents itself in writing and how it sounds in conversation.

How do I choose a TTS voice for my business?

Start with the use case, not the voice catalog. A customer support agent needs a calm, measured voice. An outbound sales agent needs energy and pace. A compliance line needs clarity and authority. Then test voices against real call scenarios — not studio demos — with your actual scripts, product names, and the background noise your callers will have. Evaluate accuracy on numbers, names, and domain terms, not just how pleasant the voice sounds.

Should I use a cloned voice or a stock voice?

Stock voices work well for most use cases and are faster to deploy. Voice cloning makes sense when you need continuity with an existing brand voice — for example, if your company already has a recognized spokesperson or a voice customers associate with your brand. Cartesia offers instant voice cloning from 5 seconds of audio and professional voice cloning from 30+ minutes for higher fidelity.

What voice characteristics matter most for voice agents?

The three that matter most are intelligibility (can callers understand numbers, names, and terms correctly), emotional register (does the voice match the tone of the interaction), and latency (does the voice model respond fast enough for natural conversation). A voice that sounds beautiful but mispronounces your product name or takes two seconds to start speaking will hurt the customer experience more than it helps.

Can I change my brand voice later?

Yes. With Cartesia's voice library, switching voices is a configuration change, not a re-recording project. You can test different voices in the playground, A/B them with real callers, and swap them in production without changing your agent's code. That said, frequent voice changes can confuse returning callers, so treat the initial choice seriously.