Generic restaurant voice AI demos are almost always built on a menu of burgers, pizzas, and wings. Those menus are phonetically easy, have three modifiers, and never require the caller to switch languages mid-sentence. An Indian or South-Asian restaurant phone line is a materially harder problem — and it breaks in ways a generic demo will never show you.
Here's what actually goes wrong, and the specific things worth testing before you buy anything.
Four things that break on a South-Asian menu
1. Speech-to-text mangles the dish names
Every voice system starts by converting the caller's speech to text. That step is trained overwhelmingly on English, and it fumbles predictably: paneer becomes “pioneer,” saag becomes “sag,” keema becomes “cream a,” biryanicomes through half a dozen ways depending on the caller's accent. If the system only matches what it literally heard against your menu text, it finds nothing and asks the caller to repeat — which is exactly the moment people hang up and call the place down the street.
The fix isn't a better microphone; it's aliases. Each menu item needs the common mis-hearings and spelling variants attached to it, so “chicken tikka masala,” “tika masala,” and “TKM” all resolve to the same item. That is configuration work on your specific menu, and any vendor who says it isn't needed hasn't run a South-Asian menu through their system.
2. Callers code-switch mid-sentence
Real Hinglish isn't “Hindi calls” and “English calls” in separate buckets. It's “bhaiya, ek garlic naan aur do butter chicken, thoda spicy karna” — one sentence, both languages, numbers and dish names in whichever comes to mind first. Systems that pick a language at the start of the call and lock to it will handle the first sentence and then degrade, or answer in the wrong language entirely.
What you want is language handling at the turnlevel, responding in whatever the caller just used, with no “press 1 for Hindi” gate in front. Our page for Indian & South-Asian restaurants covers how Depla approaches this, and the earlier post on serving Hindi-speaking customers makes the business case for offering it at all.
3. The dietary question comes before the order
On most menus, modifiers are an afterthought at the end of the order. On a South-Asian menu, the qualifying question usually comes first: is it veg, is there egg in it, do you do Jain, is the paneer fried, how hot is “medium” actually. A system that can only take orders and not answer questions will fail the first thirty seconds of a large share of your calls.
Test this directly: ask a dietary question before ordering anything, and see whether the system answers from your actual menu or improvises. Improvising here isn't a cute quirk — telling a caller something is vegetarian when it isn't is a real problem.
4. Spice level and portion aren't optional
“Two butter chicken” is not a complete ticket. Spice level, and often half/full or family-size portion, has to be captured per item or the kitchen ends up calling the customer back. This is a menu-modelling question more than an AI question: those choices have to exist as structured modifier groups on each item, not as free text the model hopefully remembers. See how AI order taking handles items, modifiers, and read-back.
How to actually test a vendor
Don't evaluate on their demo menu — evaluate on yours. A short, specific script:
- Order your three hardest dish names, said at normal speed, ideally by someone with the accent your regulars have.
- Switch languages mid-sentence and see whether the reply follows you or snaps back to English.
- Ask a veg / Jain / allergen question before ordering, and check the answer against your own menu.
- Change your mind halfway— “actually make that one spicy, and drop the naan” — then listen to the read-back.
- Confirm the read-back is complete: every item, every modifier, the total, and the pickup time. A confident-sounding summary that omits a modifier is worse than no summary.
If a vendor can only show you a scripted recording rather than let you do this live on your own menu, that itself is the finding.
What this means for operators
The honest position: none of this is magic, and the quality difference between systems on a South-Asian menu comes down to unglamorous setup — aliases on the dish names, modifier groups modelled properly, and language handled per turn rather than per call. A vendor that treats your menu as a configuration problem will do well on your phone. One that treats it as a rounding error on a burger menu won't.
You can put the specific tests above to Depla right now on your own menu — try the live demo, no phone number or signup required — and the plans it runs on are on the pricing page.