Accents, background noise and bad lines
This is the question people are too polite to ask in a sales conversation and the one they should ask first. Speech recognition is very good and it is not perfect, and what a system does at the edge of its ability matters more than how it performs in the middle.
What causes most trouble
A building site, a busy road, a pub. This affects people too — you would ask them to repeat it.
The most common real-world condition and one of the least forgiving.
Not accents so much as the fact that these are unpredictable strings where context cannot help. This is where errors concentrate.
Both reduce accuracy, and both are common when somebody is stressed.
Accents, honestly
Strong regional accents are handled far better than they were a few years ago, and for ordinary conversation this is mostly a solved problem across the accents our customers actually have. It is less solved for unusual proper nouns spoken quickly in an unfamiliar accent over a poor line — which is exactly the situation where a human receptionist also says "sorry, can you spell that for me?"
The behaviour that matters
Asking again. A system that confidently writes down the wrong name and the wrong number has produced something worse than no booking, because you will not know it is wrong until you ring an unobtainable number. Checking back — repeating a number, asking for a spelling — looks like a limitation and is in fact the feature. The failure mode you want is "sorry, could you say that again", not silent confidence.
What you can do about it
Two things help more than anything else. Have the caller’s number captured automatically rather than dictated, so the most error-prone item is not spoken at all. And where a call is genuinely unintelligible, it should reach a person rather than persist — no system should spend two minutes failing politely at somebody standing next to a cement mixer.
Frequently asked
Generally yes for ordinary conversation — this is much better than it was a few years ago. It is harder for unusual names spoken quickly over a poor line, which is where a human receptionist also asks you to spell it.
Building sites, busy roads and car speakerphone are the hardest conditions. They affect people on the phone too.
It should ask again. Confidently writing down a wrong name and number is worse than no booking, because you find out when you ring an unobtainable number.
Capturing the caller’s number automatically rather than having it dictated, so the most error-prone item is never spoken.
It should reach a person. No system should spend two minutes failing politely at someone standing next to a cement mixer.