If AI phone calls are starting to sound noticeably more natural, there's a specific reason — and it isn't just "the models got better." The plumbing underneath is changing shape.
The classic way: a relay race through text
For most of the last decade, an AI voice call worked as a cascade of separate systems. Your speech was transcribed to text; a language model read the text and wrote a reply; that reply was converted back into speech. Four handovers, each adding delay — and, crucially, each throwing information away.
The relay works, and it powers a great deal of what's live today. But it has two built-in costs. Every handover adds latency (see the half-second pause). And the very first step — turning speech into text — discards everything that wasn't words: your tone, your hesitation, the rising pitch that meant you weren't finished. The model at the centre only ever sees a flat transcript. Whatever warmth comes out at the end is essentially reconstructed from a guess.
The newer way: keep audio as audio
Speech-to-speech models collapse that relay. Instead of transcribing to text and back, they take audio in and produce audio out through a single system that never fully leaves the sound domain. Two things follow:
- It sounds more human, because tone, emphasis and timing aren't thrown away and rebuilt — they carry through.
- It can be faster and more fluid, with fewer handovers to add delay, which also makes natural interruptions easier to handle.
That's the real reason some AI calls have crossed from "clearly a robot" to "wait, was that a person?"
The honest trade-offs
Newer doesn't mean strictly better for every job. Keeping everything in the audio domain makes the system harder to constrain. For an open, friendly receptionist chat, expressiveness is the whole point. But for a regulated survey, where every question must be asked in exact, approved wording and the same order every time, the tight control of the text-based approach can be an asset, not a limitation. The right architecture depends on whether you're optimising for warmth or for precision — and serious systems increasingly blend the two.
The takeaway isn't that one approach won. It's that "why does this call sound so much better than the last one?" now has a concrete answer — and it's about plumbing, not magic.
🎧 Hear the tech explained in plain English in our audio briefing — listen here.
Related: The half-second pause: latency on an AI call · How an AI phone call works, end to end