New AI voice models for restaurant phone ordering, April 2026: what actually survives a Friday rush

Three released between December 2025 and April 2026 are worth attention. OpenAI gpt-realtime is now generally available with two new voices (Cedar and Marin) and a 20% price drop versus gpt-4o-realtime-preview. Cartesia Sonic-3 is the latency leader, with time-to-first-byte under 100 milliseconds and a Sonic Turbo variant at 40 ms. ElevenLabs opened the public API for Eleven v3, with audio tags ([excited], [whispers]), multi-speaker dialogue, and 70+ languages. The rank order in studio benchmarks does not match the rank order on a real restaurant phone line.

Restaurant phone calls arrive over the public switched telephone network as 8 kHz G.711 mono. Vendor demo audio is 22 or 24 kHz studio capture. The frequency content above 3.4 kHz is gone before the model ever sees it. On top of that, the audio is mixed with kitchen ambient noise (hood vents, ticket printers, dishwashers), the caller is often in a moving car or on a Bluetooth headset, and there is almost always a side conversation in the background. A model that produces flawless prosody on a 24 kHz read-through of a paragraph can still mis-hear 'mild' as 'wild' on a $24 large two-topping order.

Four axes that map to revenue, not to vendor leaderboards. (1) Multichannel ASR accuracy on 8 kHz mono with the customer and AI on separate channels (Deepgram nova-3 multichannel is the current default). (2) Barge-in latency: the time between the caller starting to speak and the AI yielding the floor, measured under simulated kitchen noise. (3) Cuisine-name and modifier accuracy on the actual restaurant's menu (half-and-half pizzas, spice levels, protein subs, custom sushi rolls). (4) Round-trip POS commit latency: from caller saying 'that is everything' to a fired ticket on Clover, Square, Toast, NCR Aloha, or Revel. The reference for all four is a 102.36 second stereo recording at /public/audio/dennys-order.mp3 that ships with this site.

Strong: instruction-following on multi-turn modifications ('half pepperoni half mushroom, light cheese on the mushroom side, no garlic'), tool-call precision against a structured menu schema, and the new MCP support reduces glue code on the agent side. Weak: ambient noise sensitivity is higher than Cartesia in our barge-in tests, and the speech-to-speech path can over-confirm when the caller is curt. The new Cedar and Marin voices sound notably more natural than the older Alloy/Sage but the price is still $32 / 1M audio input tokens and $64 / 1M audio output tokens, which adds up at one large pizza order per minute on a Friday.

Strong: time-to-first-byte under 100 ms is the difference between a phone call that feels like a real conversation and one where the caller talks over the agent. Sonic-3 also adds fine-grained control on volume, speed, and emotion, and the new audio-tag-style markup lets you bias intonation per item read-back. Weak: it is a TTS model, not a speech-to-speech model. You still need a separate ASR (we use Deepgram nova-3 multichannel) and an LLM brain. That is fine for a restaurant pipeline because we want each component swappable, but it is not a one-model drop-in.

Strong: cuisine-name pronunciation across languages. 'Dosa', 'pho', 'al pastor', 'crudo', 'okonomiyaki' are pronounced correctly without per-word phoneme overrides. Multi-speaker dialogue is interesting on hold-music style scripts, less relevant on a one-on-one phone call. Weak: latency is higher than Sonic-3, and the audio-tag DSL is expressive but easy to over-tune. We use Eleven v3 selectively for restaurants whose menu sits in two or more languages and whose customer base reads it in either.

We rebenchmarked all three. The default ASR remains Deepgram nova-3 multichannel because per-channel word timestamps are load-bearing in our monitoring pipeline (the same pipeline that produces the 60 Hz amplitude envelope on the home page). The default speech-to-speech path is being moved to gpt-realtime for restaurants on annual plans because the tool-call precision improvement on multi-modifier orders is real. Sonic-3 is the default TTS for the menu read-back path, where the latency improvement is most audible to callers. Eleven v3 is offered as an option for multilingual menus.

No. The API is a substrate, not a product. To turn it into something a restaurant can answer Friday's calls with, you need: a menu ontology mapped to your POS item IDs (dish names, modifiers, allergens, half-and-half rules), a barge-in policy tuned for kitchen noise, a confirmation-phrase library for your cuisine, a transfer policy for edge cases, payment handling for delivery, a structured per-call record so a manager can audit, and a same-day onboarding loop. PieLine builds all of that on top of these models so the operator does not have to.

They do not, on PieLine. Standard pricing is $350/month for up to 1,000 calls, $0.50 per call after that, with a money-back guarantee for the first month. Underlying model cost is our problem, not the operator's. Per-minute pricing from some competitors is unpredictable for restaurants and tends to spike exactly when the line is busiest. We absorb the model swap so the operator's bill does not change when OpenAI or Cartesia ships a new SKU.

Both are real and both are good signals for the category. Burger King began piloting an OpenAI-based voice assistant called 'Patty' inside employees' headsets across 500 restaurants. ConverseNow and Deliverect announced an integration on April 16, 2026 to flow phone and drive-thru voice orders into Deliverect's unified order management. PieLine sits in a different lane: independent restaurants and mid-market chains where same-day onboarding and per-call pricing matter more than enterprise procurement. The model wave lifts every boat, but the buyer profile is different.