Why does voice AI keep getting addresses and part numbers wrong?
Because speech recognition leans on what is likely to be said, and addresses, surnames and model numbers are deliberately arbitrary. There is no linguistic context to lean on, so the recognizer falls back to the closest common word. Fixes are structural: bias the recognizer toward your real service area and equipment list, confirm the value back to the caller, and move genuinely high-stakes strings to keypad entry or a texted link.
Why proper nouns are the hard case
Recognition systems combine what the audio sounds like with what a sentence is likely to contain. That second half is enormously helpful for ordinary speech and actively harmful for identifiers. "Sixty-fourth" and "sixty-fourth street" behave well. A street name that exists in three towns, a surname with an uncommon spelling, or a compressor model number carries no probability mass at all, so the recognizer substitutes something that sounds close and occurs often.
This is not a bug that gets fixed by a better model. It is a property of the problem. The engineering answer is to give the system the arbitrary information in advance.
Bias the vocabulary to your actual business
A home services company does not serve every street in the country. It serves a defined set of zip codes with a knowable list of street names, and it works on a knowable list of equipment brands and model families. Feeding that list into the recognizer as a bias vocabulary changes the odds dramatically for exactly the words that were failing.
The same trick applies to your job types, your technician names and your membership plan names. Anything that is a closed set in your business should be a closed set in the recognizer. Pulling those lists out of ServiceTitan or Jobber and keeping them in sync is ordinary integration work, and it is the highest-leverage accuracy fix available.
Confirm the way people actually confirm
- Read it back in chunks. Numbers grouped the way people say them, not digit by digit, unless the caller is spelling.
- Confirm once, not three times. Repeated confirmation is the fastest way to make a caller ask for a human.
- Offer a spelling fallback only after one failure, not preemptively.
- Text the confirmation. A message with the address and time lets the caller correct it visually, which beats any voice loop.
Diagnose before you tune
Errors that cluster tell you the cause. If the same three street names fail across many calls, that is a vocabulary problem and the fix is a list. If failures are scattered evenly across all values, that is an audio problem and the fix is upstream of the model. If they concentrate in one time window or one carrier, look at the channel.
Pulling that distribution requires transcripts joined to what was ultimately booked, which is what call analysis exists to provide. Without it you are tuning on impressions.
Topics: speech recognition · addresses · data quality · confirmation design
Have a version of this question about your own business?
The useful answer usually depends on which systems you run and how they're connected. That's a conversation, not a blog post.