How accurate is the transcript?

Speech recognition works on probability. It picks the most likely sequence of words given the audio and what usually follows what.

That is why it handles a normal sentence well and a surname badly. There is no context to make an unusual name more likely than a common word that sounds like it.

Numbers are the other weak spot, because a misheard digit is still a plausible digit and nothing flags it as wrong.

Names and addresses. The fix is to have the assistant read them back and spell them where it matters, which costs a few seconds and saves a wrong callback.

Phone numbers. Read back in groups and confirm. This is the single highest-value confirmation on any inbound call, because everything else depends on being able to call them.

Background noise. A caller in a van on a motorway is a genuinely hard case, and no vendor's headline accuracy figure was measured there.

Overlapping speech. When the caller and the assistant talk at once the transcript gets confused, which is one reason turn-taking behavior matters as much as recognition quality.

Say you run an appliance repair business. A caller on a cell phone in her kitchen, with the dishwasher grinding behind her, says: "It's Deirdre Kowalczyk, I'm at 2260 Goethe Street." The transcript reads "Deirdre Koval check" and "2260 Gurty Street." Everything else in the call, the fault, the brand, the preferred day, came through word for word.

That's the typical pattern. Ordinary sentences survive and proper nouns get bent into familiar-sounding words. It's also recoverable here. Her number arrived with the call as caller ID, so nothing depended on hearing the digits. There's no Gurty Street in your town, and a map search for 2260 turns up Goethe in seconds. The surname gets fixed when your tech phones ahead.

The risky error is one that looks fine. "2216" heard as "2260" is a real address on the same street, and nothing in the transcript hints that it's wrong.

Once a month, pull twenty calls that turned into real jobs or appointments. For each, compare three transcript fields against what you later learned was true: the name, the address, and any callback number that was spoken instead of captured from caller ID. Tally the misses per field. You'll end up with a small table that says something like names 4 of 20, addresses 2 of 20, numbers 0 of 20.

That table tells you where to spend effort. Suppose addresses are the problem. The assistant reads details back to confirm them, so make sure the address is one of the details it confirms, then test that flow with your trickiest street names. If recording is on, listen to the audio for each miss, because you'll often hear the cause: a speakerphone, a car's Bluetooth, or a job site compressor.

Expect the audit to differ by caller type. Landline callers in a quiet house produce cleaner text than contractors phoning from a roof. A business whose callers are mostly on job sites should budget more human checking of names and addresses than a business whose callers are mostly at desks.

Every call arrives as a full transcript plus a short summary, and the transcript is the record. The summary is for skimming, not for acting on.

You can have the assistant confirm the name and number back before ending the call, which turns a recognition error into a correction while the caller is still there.

Reading transcripts in the first week is how you find your own error pattern. A practice with a lot of unusual surnames has a different problem from a plumber taking calls from building sites.