How to Choose a Voice Model for Your AI Receptionist: Cascaded vs Speech-to-Speech vs Audio-Native

Most buyers never hear about it because vendors sell "an AI receptionist," not a pipeline. We wrote about the three things that make a receptionist sound human — speed, understanding, voice — and this is the layer underneath that decides all three. This summer the model layer split visibly — PolyAI shipped a new architecture, Bland shipped a new speech engine, Retell added OpenAI's realtime model, Genesys bolted on Deepgram and is adding ElevenLabs — and the differences are now large enough that you should ask.

The three architectures

1. Cascaded

The traditional way. Three separate models in a chain:

caller's audio → speech-to-text → language model → text-to-speech → caller hears reply

The speech recogniser writes down what it thinks you said. The language model reads that transcript and decides what to say. A voice model reads the reply out loud. Each step is a specialist and each can be swapped: better recogniser, different LLM, nicer voice.

What the language model never gets is the audio. It only sees the transcript's single best guess. If the recogniser wrote "Tuesday," the LLM has no way to know the caller might have said "Thursday," no sense of tone, no idea there's a lorry reversing in the background. A separate small model ("voice activity detection") decides when the caller has stopped talking, usually by measuring silence, which is why cascaded agents either interrupt you mid-thought or leave awkward gaps. Tuning that trade-off is a lot of what vendors' engineers do all day.

2. Speech-to-speech

One model does everything, end to end:

caller's audio → one big model → audio reply

OpenAI's GPT Realtime and Google's Gemini Live are the well-known examples. The model hears you and speaks back, so in principle it can react to tone, hesitation and background noise, and turn-taking feels more natural. The cost is control. The voice is baked into the model with limited options, so it may pronounce your brand name wrong and there's no dial to fix it. It's also expensive to run: some versions keep a GPU pinned for the entire call, and OpenAI's per-minute pricing was in the $0.06 to $0.10 range as of mid-2026 before anything else in the stack.¹

There's also an open question over how end-to-end some of these models really are. PolyAI's research team published benchmarks in July claiming GPT Realtime 2.1 scores on par with cascaded systems on audio-aware tests, and suggested it may be "a well-engineered cascaded system" underneath.² That's a competitor's claim about a rival, so weigh it accordingly, but it's a testable one: they suggest asking it whether Americans say "X Y Zed."

3. Audio-native (input side)

The newest approach, and the one that's interesting for phone work:

caller's audio → one model that hears the audio and decides when to speak and what to say → separate text-to-speech → reply

The listening half is fused: turn-taking, speech recognition and the reasoning about what to do happen in one model that has the raw audio and the full conversation context. The speaking half stays separate, so you keep full control of the voice, pronunciation and cost.

PolyAI's Dialog-RSN-1, released 30 July, is the clearest example. Their claims: sub-300ms response, an 11% relative lift in containment at a restaurant group, a 37% cut in latency at an insurer, and turn-taking that can be steered with plain instructions like "wait for all six digits of the customer code before responding" or "if the caller thinks they're talking to voicemail, interrupt them."² It's English-only for now and currently available to existing PolyAI customers with an early-access list for everyone else.

The reason this matters for a receptionist rather than a research lab: the failures callers notice most — being talked over, being asked to repeat a postcode, having a name misheard — are turn-taking and recognition failures, and those are exactly what this architecture moves into a model that can reason about them.

What each one means for the person on the phone

Two honest notes on that table. "Best in principle" is doing work: a well-engineered cascade with a good recogniser beats a badly configured anything. And the only row you can verify yourself is the caller experience one, which is why the questions at the end are about test calls, not architecture diagrams.

What the vendors shipped this summer

You don't need to track this, but it explains why the sales pitches changed.

• PolyAI — Dialog-RSN-1 (30 Jul), the audio-native input model above.²
• Bland — Bland Speech v3 (4 Aug), which it calls a "Human Speech Engine," on the output side; and a day earlier, "adaptive resumption," which measures each caller's pace and adjusts how long the agent waits before replying, with the manual setting removed entirely.³ That's a cascaded vendor attacking the turn-taking problem from the other direction.
• Retell — added GPT-5.6 and GPT Realtime 2.1 to its agent builder (24 Aug), plus a fix for iOS and Google call screening being mistaken for voicemail, and Soniox's Spanish recogniser on every call.⁴ Retell lets you pick the model per agent, which is the cascaded model's advantage in one line.
• Genesys — added Deepgram real-time speech recognition to its Agentic Virtual Agent at Xperience (1–3 Sep), with ElevenLabs voices due by end of October and its own end-of-turn detection the quarter after.⁵ The biggest contact-centre platform in the world is assembling a cascade from best-of-breed parts.

The direction is clear enough: everyone is trying to fix turn-taking and recognition, some by fusing models and some by tuning the joints between them. For a clinic or garage buying a packaged receptionist, the architecture is the vendor's problem. Whether it works on your callers is yours.

Where this leaves a clinic or a garage

If you're buying a packaged AI receptionist, you mostly won't get to choose the architecture. What you can do is know which one you're getting and test for its weaknesses.

A cascaded product will be cheaper and more configurable, and its weak spot is turn-taking and accents. Test it with a caller who pauses mid-sentence and with the strongest regional accent among your actual patients.

A speech-to-speech product will sound the most natural in the demo, and its weak spots are your brand name, cost at volume and the fact that you're really buying OpenAI or Google. Ask it to say your clinic's name and your lead physio's surname.

An audio-native product is the newest and, if the claims hold, the best fit for phone work, and its weak spots are that it's new, currently English-only in PolyAI's case, and tied to one vendor's model. Ask what happens to non-English callers.

Apex AI's position

We build cascaded, and we build it the expensive way. Our receptionists run a frontier language model (Anthropic's Claude) on real-time voice infrastructure (Vapi), with premium British voices on the output side; we publish that on our GDPR statement because a clinic is entitled to know who processes its calls. The reason for cascaded isn't cost. It's control: we can make the agent pronounce your clinic's name and your lead physio's surname correctly, swap the recogniser when a better one ships, and pick the model per client rather than being tied to whichever voice OpenAI or Google baked in this quarter.

The trade-off is that turn-taking and recognition are the joints in a cascade, and that's where we spend the engineering time: keeping pickup under eight seconds and the reply gap close to conversational, tuning interruption handling, and testing every build on accents and noisy lines before it goes live. We'll move to an audio-native input model when one is available to us in production with the same control over the voice, because on paper that's the right architecture for phone work. It isn't there yet for a UK small business, and we'd rather tell you that than dress up the roadmap.

The proof is on the outcome side rather than a benchmark chart: during an NHS referral surge, our receptionist answered 400 calls for Physio and Health Matters in a month, missed none, and booked 101 appointments. That's a cascade doing its job on real UK callers.

Five questions to ask any AI receptionist vendor about their voice stack

1. Which speech recogniser, which language model, which voice — and can any of them be changed? If they won't say, that's an answer.
2. How does it decide the caller has finished speaking? "Silence threshold" means cascaded and you should test interruptions. "The model decides" means ask for the evidence.
3. Can you make it pronounce our name correctly? Then make them do it on the call.
4. What's the response time on a real phone call, not a demo on a laptop? Ask for a number and then ring it from a mobile in a car.
5. What happens with a strong accent or a noisy line? Don't accept a description. Ring it from the workshop with the compressor running and a colleague from Glasgow on the line.

The architecture matters. The test call matters more.

Apex AI builds AI receptionists for clinics, physios, MOT centres, trades and care homes. Ring ours, interrupt it, give it your worst accent and see. Test our agent.

Sources