Nobody at a Counter Talks Like a Database

I added voice ordering to Bakery Buddy this week. You tap a mic, a customer orders out loud like they're standing at the counter, and the ticket fills itself in as they talk.

Everyone assumes the hard part is the speech recognition. It isn't. That part works. The hard part is what happens after the words are correct.

The gap nobody warns you about

Here is a real order, exactly as someone says it:

"I need 45 cookies, the rainbow ones, 4 fruit tarts, 3 napoleons, maybe 10 cannolis."

Now look at what your menu actually stores. A catalog says Round 6". A customer says "six inch round." Those two strings share almost no characters. The quantity shows up before the item, not after. "The rainbow ones" is not what the menu calls them. People correct themselves mid sentence. There is no punctuation anywhere in speech.

The model hears all of it perfectly. Then it has to reconcile how a person talks with how your data is shaped, and those are two different languages.

What I actually had to build

The fix wasn't a smarter model. It was three unglamorous normalizations:

• Spoken number words map to digits, so "six" and 6 reconcile
• Inch marks and the spoken word "inch" fold away, so 6", six inch, and 6 all collapse to one token
• Matching scores on how much of the catalog name got covered, not how much of the spoken phrase did. That way a terse entry like Round 6" still matches inside a long phrase like "six inch round vanilla"

Then I stopped inventing a menu to test against and seeded the catalog from the top 25 items my wife Lindsay has actually sold at her cake studio. Testing against real orders surfaced problems a made-up menu never would.

The lesson I keep relearning

The typing got cheap. An AI agent can write the matching function in seconds. But it can't tell you that bakers say "six inch" and the database says 6", or that quantities come first, or that a wrong cake is worse than a slow one. That judgment is the whole job now.

Nobody at a counter talks like a database. Closing that gap is the work.