A brand voice is easy to describe in a creative brief: warm, clear, calm, and professional. Those words are harder to turn into a dependable experience on a noisy phone line, especially when the agent needs to read a date, explain a delay, or recover from a misunderstanding.
On October 16, 2026, OpenAI’s API changelog announced prompt-to-voice availability for API customers. Its custom voices documentation describes generating a voice from a text description for GPT-Live sessions. That makes voice exploration more flexible; it does not remove the need to test whether customers can understand and use the result.
Give the voice a job before judging its personality. Can listeners catch the appointment time? Can they follow a clarification? A short, repeatable test will answer more than a roomful of preferences.
Write a brief around the job
“Sound friendly” is too broad to evaluate. Write down what the voice must help a caller do. A receptionist may need to ask concise questions, communicate uncertainty without sounding evasive, and make the next step easy to understand.
List constraints alongside the desired tone. Specify expected languages, likely line quality, common names, and the kinds of numbers the agent will repeat. A voice that is attractive in a short welcome may be difficult to follow through a long address or reference number.
Do not use a recognisable person’s identity as a shortcut for the brief. Describe communication characteristics and choose an authorised voice workflow suitable for the project.
Build a script that exposes weaknesses
Use the same test script across candidates. Include ordinary greetings, short answers, names, dates, contact details, and a polite clarification after an error. Add phrases with similar-sounding words and a sentence that should be spoken slowly.
For multilingual work, test the actual language combinations and vocabulary expected in your business. A filter or model language list cannot tell you how local place names or code-switching will sound in context.
A useful comparison asks listeners to repeat back the important information. Preference alone is not enough. Someone may like a voice while consistently mishearing the appointment time it reads.
Evaluate a conversation, not just an audio sample
An isolated sample misses turn-taking and tool behaviour. Place candidate voices inside the same conversation flow and run the same scenarios.
Ask whether the agent leaves space for the caller, whether it handles a correction naturally, and whether it signals uncertainty clearly. Review moments when the caller must wait for a system response. A calm voice should not create the impression that an unconfirmed action is already complete.
Keep prompts stable during the comparison. Changing instructions, model settings, and voice simultaneously makes it difficult to determine what caused a better result.
Include the delivery channel
A browser demo and a phone call are different listening conditions. Test through the intended connection, devices, and environments before rollout. Include a mobile speaker, a headset, and a realistic level of background noise.
Measure comprehension, repeat requests, successful task completion, and caller effort. Record which version of the voice and configuration produced each result. If a provider changes behaviour, a reference set gives your team something concrete to compare.
Agree on failure conditions in advance. Repeatedly misunderstood numbers or an unsuitable accent should be a reason to revisit the choice, even if internal reviewers like the sound.
Turn a voice choice into an operating decision
Select an owner for voice configuration and a process for approving changes. Keep a known-good configuration available for rollback. Re-run a compact test set after changes to prompts, speech providers, or call routing.
The voice should fit the task rather than distract from it. A customer calling to resolve a problem usually needs a clear answer and a reliable next step more than a distinctive performance.
Explore the InferScale approach
InferScale Voice features include supported provider configuration, voice exploration, and browser testing. We confirm provider-specific capabilities during deployment planning.
You can try the voice demo to understand the interaction, then discuss your languages and call types through our custom voice-agent service. A discovery call can turn a broad brand brief into an evaluation plan with useful acceptance criteria.
Sources checked October 18, 2026: OpenAI API changelog and custom voices documentation.