An AI voice for your phone greeting: who checks what it says?

Who listens when an AI speaks for your company?

For most people who use one: nobody. You listen to the first three recordings, like them and switch the voice live. After that only one person is still listening — the customer.

In an AI-run back office a different rule applies: whatever a machine does, a second machine checks before it goes out. This LAB shows the rule on an example you can hear — our blog posts read aloud.

The setup

Every sentence is spoken on its own. Then a second model — the open speech recogniser Whisper — turns the recording back into text, and a program compares that text with the script. If it does not match, the sentence is spoken again. At most ten times; the best recording wins.

Each of these steps is written to a log: the sentence, the number of attempts, the deviation from the script and the words the checker could not find again.

The numbers

Period 28 August to 3 October 2026, straight from the log:

• 7,708 sentences checked, 3,786 German and 3,922 English — together 9.4 hours of audio.
• 79.0 per cent are right on the first attempt. Every fifth sentence was sent back by the checker before anyone heard it.
• All sentences together took 12,457 attempts, 1.62 per sentence on average — spoken and checked without a human having to step in.
• For 77.7 per cent of the sentences, the text heard back deviates from the script by at most five per cent of the characters in the end.

Three findings

Short sentences need more checking than long ones. Sentences of three words or fewer are right on the first attempt 66.4 per cent of the time, sentences of twenty words or more 92.1 per cent. A single word without a sentence melody slips more easily than a whole thought.

A checker needs a list of names. What the checker misses most often is a proper name — our company name. It is spoken correctly; the checker just spells it differently. Without a list of its own names it raises false alarms, and a checker that keeps raising false alarms is believed by nobody after a week.

Every sentence is measured, not a sample. Whoever only listens to the recordings they happen to click knows nothing about the others. The checker hears all of them — including the nine hours no human would listen to in one go.

What this means for your office

Phone announcements, read-aloud functions, voice bots on the customer line: as soon as a machine speaks for your company, the same question comes up. Not «does it sound good?», but «who listens when I am not listening?». In the AI back office this applies to every task: what runs automatically is checked automatically — and a human only sees what the check could not settle.

Who listens at your company when your phone announcement plays?