Tech Tracker: Inside the Next Generation of Restaurant Voice AI

I am sitting in a side room off the San Jose Convention Center floor, the GTC keynote chatter still bouncing around the lobby, and I am not reading the Yum/Nvidia release. I am reading NRN’s Tech Tracker, which dropped this morning under the headline “How the next generation of voice AI is getting smarter.” It is the most important restaurant-tech document of the day, and almost nobody at GTC has clicked it yet.

Here is what it says, and here is why the competitive map you had in your head on February 18 is wrong on March 18.

The 30-day reset

Read the Tracker straight through and the pattern is impossible to miss. SoundHound has unveiled what it calls its NextGen platform — a fully automated voice-AI stack with no human-in-the-loop, designed to run drive-thru, call-to-order, text-to-order, scan-to-order, and in-vehicle ordering off a single brain. Foreva.AI has launched as a new entrant aimed squarely at the same surface area. Incept is shipping accuracy gains it is willing to talk about on the record. Kea has rolled out a self-service tier — the same Kea whose CEO Adam Ahmad tells the Tracker that what used to take “three to six months” to stand up can now happen “under an hour.”

Four vendor moves. One 30-day window. A category that spent 2024 looking like a slow grind of pilot programs has reset itself while operators were still arguing about whether voice AI was real.

The lazy read is “everyone announced stuff, nothing changed.” The lazy read is wrong. What changed is the axis the category competes on. Through 2024, voice-AI vendors competed on order accuracy benchmarks and logo counts — who had Carl’s, who had White Castle, who had Bojangles. After today’s Tracker, the segmentation is different, and operators who do not see it are going to buy the wrong product.

The new segmentation: HITL % and deployment mode

The two variables that now define the map are how much human-in-the-loop the system requires, and how many ordering surfaces it covers from a single deployment. Plot the four vendors in the Tracker on those two axes and the category sorts itself.

SoundHound NextGen is the corner play: zero HITL, multi-surface, one platform spanning drive-thru, phone, text, scan, and in-vehicle. That is a specific bet — that the LLM stack is now reliable enough that the call-center fallback is dead weight rather than a safety net. It is also a bet that operators want one vendor for every voice surface, not five.

Kea’s self-service tier is the opposite corner of the same insight. “Three to six months down to under an hour” is not a marketing line — it is a statement that the deployment cost has collapsed. When the deployment cost collapses, the buyer changes. You stop selling to the CIO at a 2,000-unit enterprise and start selling to the franchisee with twelve stores who wants voice on the phone by Friday. That is a different motion, a different price point, and a different competitive set.

Incept and Foreva.AI sit between those poles — Incept doubling down on accuracy as the wedge, Foreva.AI new enough that we will not know its real positioning for another quarter. But the corners are now claimed, and the middle of the map is where the squeeze is going to happen first.

The thing that should worry any operator reading this is that the voice-AI pilots they signed in 2024 were probably scoped against the old map. The vendor you chose because it had the best drive-thru accuracy may be the wrong vendor if you also want phone and text from the same brain. The vendor you chose because it had a managed-service HITL team may be selling against its own product in six months.

What the Tracker does not say

NRN’s Tech Tracker is a useful document because it puts four vendor moves in one frame. It is not a complete document, because it does not name the second-order effects, so let me name three.

First, the HITL layer is now a liability for the vendors who built their pricing around it. If SoundHound is willing to ship zero-HITL and Kea is willing to ship self-service, the vendors quietly billing per-call human review minutes have a margin problem that will become a pricing problem inside two quarters.

Second, the in-vehicle surface in SoundHound’s NextGen pitch is the tell. Voice AI is no longer a drive-thru category — it is an ordering-everywhere category, and the OEM relationships SoundHound has accumulated are the moat the other three do not have. Watch for Foreva.AI to either announce an automotive partner or get bought.

Third, the “under an hour” deployment claim from Kea reframes the whole time-to-value conversation we had about back-of-house AI last quarter. If deployment really is an hour, then voice AI is the first restaurant-AI category where the proof-of-concept and the production deployment are the same event. That is going to make every other AI vendor’s six-month pilot look slow by comparison.

The Yum/Nvidia news from this morning will get the headlines today. The Tech Tracker is the document that will matter in 90 days.

— Maya covers restaurant tech. Tips: tips@tabletransfers.com.