The Sixth Voice

The finale of our mini-series on Asya's six voices. After four big names, the job went to an engine you have most likely never heard of — Ultravox. This dispatch is about why the winner wasn't the most famous model, about the true price tag of the decision, and about the limitations we know by name and have no intention of hiding.

First, what it actually is. Ultravox is built as a hybrid: the hearing is embedded directly inside the model — sound enters it without an intermediate transcription into text — the thinking is done by a stack built on an open language model, and the speaking is handled by a separate fast synthesizer at the output. For us this meant two things: speed — and, as it turned out to matter more, transparency. This architecture has a visible transcript: you can always break a failed call down into "what she heard" and "what she decided" — and fix precisely the link that broke.

Anyone who has read the series this far already understands why that outweighed everything. For half a year we had been debugging black boxes: the engine goes silent — and you're left guessing whether it misheard, misunderstood or froze. Here the guessing ended. Remember the detective stories from earlier dispatches — the house with the slash, the ghost of the first second? Both were solved by reading Asya's inner draft. On an engine without a visible transcript those investigations would have been impossible — we would still be shrugging.

Now the price — our chief judge. Five cents per minute, flat. We didn't take the price list on faith and measured it on live traffic: a week, one hundred thirty-seven calls — five cents it was. Against twenty-three with the previous main engine. Meanwhile, on an independent instruction-following benchmark — and you already know that is our sore spot — it scores ninety-eight percent against eighty-seven for the big-name competitors. Cheaper by multiples and more disciplined — it wasn't supposed to work out that way, but it did.

We tested it on people too: same scenarios, same day, two engines back to back. A conversation that dragged on for three or four minutes with the OpenAI backup — repeated questions, circular detours — Ultravox closed in a minute or a minute and a half: shorter, more collected, no stalls. For someone calling to say their internet is down, that is the difference between "solved quickly" and "put through the wringer."

Now the honest part — the limitations, each by name. First and foremost: digits on a narrow telephone line. An ordinary landline connection cuts sound down to eight kilohertz, and on that input long account and phone numbers get recognized with errors — not because the model is bad, but because half the information is physically absent from the sound itself. On clean audio the accuracy is practically one hundred percent; on a phone line, when the caller is identified by their number and the account card is already open, fabrication is zero out of thirty-three verified cases; but an anonymous caller dictating a long number over street noise is a lottery — and we accepted it as a known limitation, in writing and with our eyes open.

Second: early on, Asya read service words out loud — having received the system's technical mark about a call transfer, she dutifully recited it to the person as if it were part of the conversation. We fixed it with a pause and a rule. Third: the hearing produced phantom empty turns — Asya thought the caller had said something, and re-introduced herself mid-conversation. That one is fixed too. Fourth, still alive: after a transfer she sometimes keeps thinking in the old context — we're working on it. As you can see, the glass is transparent in both directions: we show not only other people's failures but the list of our own open wounds.

And the last thing worth saying plainly. Choosing an engine is not a lifelong marriage. Ultravox won the competition of May 2026 — on our scenarios, our languages and our budget. The market moves fast, we have a live backup on the bench, and if tomorrow a candidate appears that is easier to measure and more honest — the competition reopens, and you will read about it right here, in the same tone. As it happens, we already have a few complaints accumulating against the champion — but that is material not for the finale of a series, but for its continuation.

And so the mini-series about the six voices is over — the journal is not. Asya is answering calls right now, as you read this, and tomorrow she will do something else worth writing about. Until the next dispatch.