Rovers

Voice-to-text is solved.Voice-to-outcome isn’t.

Dictation puts your words on the screen. That was never the hard part. We don’t speak the way we write, and AI learned from writing: billions of typed lines, never spoken. So an AI can hear every word correctly and still do the wrong thing, because the words were the easy part. It guesses the rest. Rovers builds the layer that resolves instead of guessing.

SpokenOperating room

hold the heparin if the platelets come back low, and page haematology

ExecutedIn the order the request asked for
01Flag the heparin holdon the 06:00 draw
02Page haematology nownot gated on the result
03Note it on the chart
The other reading pages haematology only if the platelets come back low. A phrase break decided it, and a transcript has none.
  1. Rest of the sentence
  2. Words your work uses
  3. Correction you made
  4. Your systems
  5. Conditions and order
01 / Linguistic

Two words that sound identical, and only one of them fits what comes after. A current recogniser settles this without help: its decoder is a language model, and it has already read everything before the gap.

So this rung is climbed, and transcription really has got good enough to climb it. It is also the only one it reaches. Nothing above this is a recognition problem.

02 / Lexical

Identical audio again, and this time the sentence does not settle it, because both readings parse. A general model has every reason to prefer the English word over the one that belongs to your work.

Every field has a few hundred of these, and “send the Q3 sow” reads perfectly well. It has just quietly stopped being about anything real.

03 / Conversational

You said Maya, heard yourself, and fixed it mid-sentence. Writing has no way to carry that. Speech carries it for free.

A cleanup layer reads “sorry, to Dan” as noise and deletes it. The request then reads perfectly and goes to the wrong person.

04 / Situational

Which SOW. Which Dan. Not what the words mean — what they point at inside the tools this has to land in.

None of this was in the audio. No transcript can recover it, however good, because it was never said out loud.

05 / Operational

When it may run, and what has to be true first. Speech carries both in one breath. A form has no field for either.

One action, one time, one guard — and every part of it traceable back to something you actually said.

What you said

um, send the Q3 sow to Maya — sorry, to Dan — write right after the call, and only if legal have signed off

Resolved0/7of what the request needs
01 / The case for speaking

The keyboard won by default, not by fit.

Speaking is not a fallback for people who cannot type. It carries things a keyboard has no way to accept — and for most jobs it is simply the better instrument.

Same four gaps. The only question is who closes them.

Spoken · exophoriclistening

move the launch to the 14th and tell the three teams that are blocked

  • the launchHelios 2.0
  • the 14th14 March
  • the three teamsPayments, Identity, Mobile
  • blockedon the API freeze
  1. Helios 2.0 sits on 7 MarchMoved to 14 March
  2. four milestones hang off itMilestones shifted
  3. the API freeze lands firstRoadmap updated
  4. Payments, Identity, MobileThree teams notified
0:00four seconds to say. Eight more for the harness to fetch the four gaps and act on them.
Written · endophoricready

Typed out in full, it reads: Move the Helios 2.0 launch from 7 to 14 March, shift the dependent milestones with it, and update the roadmap. Tell Payments, Identity and Mobile by email — they’re the three blocked on the API freeze. Don’t tell the customer yet. Forty words, one minute and twelve seconds. The spoken request had already run at twelve seconds, by which point only the first 9 characters of 230 had been typed.

0:00forty words at forty a minute, plus the lookups. The mark is twelve seconds in — where the other lane had already finished.

Same day, same person — eleven things finished, or nineteen.

Model, not measurementNobody was timed. This is arithmetic on three published numbers and one eight-hour day, and it assumes 10 turns to hand a task to an agent, 60 seconds to leave the work and climb back in on each of them, and speech entering 3× faster than a keyboard. Move an assumption and the count moves with it. The sign does not.

One working day, 10 turns to a task finished never got to it
Typing0
Speaking0
+8more a day · +73%

Ten turns a task. Typing pays a context switch on each; speaking pays none, because you never left the work.

A switch costed at sixty seconds, not the twenty-three minutes usually quoted — that figure measures how long before you return, not what returning costs.

That is the arithmetic for a job where typing is merely slower. For most jobs it is not merely slower — there is no free hand to type with at all. Here is who that is.

Are you already talking anyway?

A consultation, a call, a deposition, a standup. The speech is already happening, so capturing it costs nobody a change of habit.

Are your hands and eyes already committed?

On a patient, inside a machine, on the wheel. When the hands are gone the keyboard is not slow. It is absent.

Does the request carry conditions and an order?

Only once legal signs off. Hold the dose unless the bloods come back low. Speech carries conditions. Forms have no field for them.

Does one sentence touch several systems?

Typing it means opening each one and restating the same request in a different vocabulary, until you have said it four times.

Nobody needs all four.

Most jobs have two, and two is enough to make the keyboard the wrong instrument for the work — not slower than speaking, wrong for it.

Who is workingAlready talkingHands committedConditions and orderSpans several systems
Clinicianin consultationThe whole consultationOn the patientDose, hold, escalate — in orderNotes, orders, messages
Field engineeron siteOn the radio, sometimesInside the machineIf it fails again, then swapWork order, parts, log
Product managerat a deskStandups and callsFreeOnly once legal signs offFour tools before lunch
Saleson a callIt is the jobWatching the customerFollow up unless they repliedThe CRM, mostly
02 / One meaning, four ways in

Say it however you think.

Understanding your language is the solved part — every current model reads hundreds of them. The systems underneath are still named in English, and a request has to reach those objects exactly, whichever way it arrived.

Typed · English

exact · endophoric

Share SOW_Northwind_Q3_v4.pdf with maya.rao@northwind.co, view only

67 characters
0 spans to resolve

Nothing to resolve, because you already did — with two lookups.

Spoken · English

pointed · exophoric

Send Maya the Q3 SOW — the one I fixed yesterday. View only.

60 characters
3 spans to resolve

Three spans that mean nothing without the week you have just had.

Spoken · Español

pointed · exophoric

Mándale su SOW del Q3, el que corregí ayer, solo lectura.

57 characters
3 spans to resolve

Three spans again — and su will not settle whose it is.

Spoken · हिन्दी

pointed · exophoric

उनको ठीक किया हुआ Q3 SOW भेज दो, सिर्फ़ देखने के लिए।

Unko theek kiya hua Q3 SOW bhej do, sirf dekhne ke liye.

53 characters
3 spans to resolve

Three spans again — and उनको will not settle how many.

One outcome

SOW_Northwind_Q3_v4.pdfmaya.rao@northwind.co

View only. Nothing else — no expiry nobody asked for, no message nobody wrote.

Spanish su marks neither the possessor’s gender nor their number, so it will not say whose SOW that is. Hindi उनको is one word for a single person addressed with respect and for several at once, and the verb agrees either way, so it will not say whether one person is about to get this or three. Neither speaker was careless. Each left to the situation what the situation makes obvious.

A layer that translates first has to pick a word and commit. A layer that resolves goes and looks — one Q3 SOW was edited yesterday, and there is one Maya on the account. Four ways in, one outcome, or it was never one interface.