Voice agent testing
By George Bâtcă, Co-founder · Security testing and evaluation
Updated 7 October 2026
A voice agent answers the phone for your company. When a call goes wrong, the customer only hears that it went wrong. Your team needs to know where: did the agent mishear, misunderstand, or say it wrong? AltaSec tests each part of a voice agent on its own and the whole call, in every language you serve, so every failure points to what you need to fix.
Three places a voice agent can fail
| Part | What it does | A typical failure |
|---|---|---|
| Hearing (speech-to-text) | Turns the caller's speech into text | A time, an amount or a name is misheard: "nine" becomes "two" |
| Understanding (the AI model) | Decides what to answer and what to do | The answer is wrong, too informal, or breaks your rules |
| Speaking (text-to-speech) | Turns the answer into speech | A number, a date or an address is read out wrong, or the voice sounds foreign to a native |
A failure in one part looks the same to the caller as a failure in another. That is why we test them separately.
How we test each part
- Hearing. Native speakers record real caller sentences, with names, dates, amounts, codes and mixed-in English words. We play them to your speech-to-text and check every word that matters.
- Understanding. We send the same caller turns as text. If the agent passes when the words are typed but fails by voice, the problem is hearing, not the model.
- Speaking. We check that what your agent says is what it meant to say: every number, date and name, the right pronunciation for the language, and a voice that sounds natural to a native speaker.
- The whole call. Simulated callers, typed and spoken, go through complete conversations: bookings, refunds, changes. We check the outcome and how the call went.
We test your own setup as it runs: your speech-to-text, your model, your voice and your settings.
Every language, checked by native speakers
Voice is where languages differ most: accents, formality, how people say numbers and dates, and the words they borrow from English. Native speakers set the standard for each language we test, and our AI judges are checked against them. Today: English, Romanian and Portuguese.
What comes next
Real-world conditions are next on our roadmap: background noise, phone lines, hold music, a second voice in the room, accents and dialects. Voice cloning and spoofing tests come later. Only what is in the "Now" column of the roadmap is part of an engagement today.
What you get
The Passport: a go / no-go verdict for every task and language, by voice and by chat, with the recording and the check behind every finding, the part that failed, a fix per finding, and a list of what we could not test. See AI agent testing and our methodology.
Frequently asked questions
Do you test our own speech-to-text and voice, or your own?
Yours. We test the setup your customers reach, as it runs, and recommend fixes for it.
What do you need from us?
A way to reach the agent (a test phone number or an endpoint) and your written authorization. No code access.
Can you tell us which part to fix?
Yes. Because we test hearing, understanding and speaking separately as well as the whole call, each finding says which part failed.