Voice is a way of using the assistant, not a separate assistant. Spoken turns land in the same conversation as typed ones, and the transcript afterwards is the same transcript.
Starting and stopping
Talk in the conversation rail starts a session. A short two-note chime tells you it is live, so you know before you start speaking rather than three words in.
The composer changes state as you go. Listening while it is your turn. Speaking while it is answering. And a quiet note while it is reading your data mid-answer, which is the state people mistake for it having stopped.
Stop ends the session. Navigating away also ends it, which is worth knowing before you wander off mid-sentence.
It has the same tools
The important thing about voice here is that it is not a reduced version.
It reads the same live data. Ask about your day out loud and it queries the same records the typed version does, through the same code.
It makes the same changes. Creating, moving and cancelling a meeting all work, through the same authority the typed version uses.
It searches the web the same way, and the sources attach to the spoken answer as a pinnable strip once the answer lands. You can keep them without touching the keyboard beyond one tap.
How confirmation works without a screen
Typed changes show a confirmation card. Spoken changes cannot, so the assistant reads the change back and waits for you to agree before doing it.
Say the change, hear it repeated, agree, and it happens. Nothing is written on the first pass.
The read-back is not a formality. Transcription mishears names, dates and times in predictable ways, and "Thursday" and "Tuesday" sound considerably more alike over a phone speaker than they do in your head.
What voice is good at
Hands busy. Driving, walking, between meetings, carrying things. This is the whole case and it is a strong one.
Short factual questions. What is next. Who is on leave. How many deals closed this month. One sentence out loud beats unlocking a phone and finding a screen.
Small changes on the move. Pushing a meeting while walking to a different one, which is the moment you actually know you need to.
Thinking out loud. Occasionally the useful thing is not the answer but having said the question, and voice is better at that than a text box.
What it is bad at
Anything with a list. A spoken list of eleven items is not information, it is noise. Ask for the count, then look at the list when you are somewhere you can.
Numbers you need to hold on to. Hearing a figure is not the same as seeing one. If you are going to quote it to somebody, read it.
Noisy places. Transcription in an open office or on a street is worse than typing, and correcting a misheard name out loud takes longer than typing the whole question would have.
Anything sensitive. You are speaking your organisation's numbers out loud in whatever room you are in. Obvious, and worth saying once.
What it costs
Voice costs more than typing, and it is worth understanding why rather than being surprised by it: a spoken exchange is charged on audio as well as text, in both directions, because the model both hears and speaks.
The practical figure is that a couple of minutes of conversation costs meaningfully more than the same exchange typed. Not enough to avoid using it when your hands are busy, which is what it is for. Enough that leaving a session open while you do something else is genuine waste, because it is listening the whole time.
Stop the session when you stop talking. That is the only cost discipline voice actually needs, and it is worth making a habit before it becomes a line on a bill.
Availability
Live voice is a feature your organisation enables rather than something every account has.
If Talk is not in your rail, it is switched off rather than broken, and an administrator can turn it on. It is worth asking, because the objection is usually that nobody has considered it rather than that somebody decided against it.
When it mishears
Three failure patterns are common enough to plan around.
Names. Anything not in a common English name list is transcribed phonetically and confidently. Colleagues' names improve as you use them; customers' names frequently do not.
Days. Tuesday and Thursday are the classic pair, and the consequence of confusing them is a meeting in the wrong place in the week.
Product vocabulary. Internal jargon, module names and abbreviations are the weakest area, because they appear nowhere in general language.
The mitigation for all three is the same and it is not clever: listen to the read-back. Every change is repeated before it happens, and that repetition is where a misheard day gets caught. Agreeing over the top of it removes the only safeguard the medium has.
For questions rather than changes there is no read-back, so a surprising answer is worth re-asking in different words before you believe it.
Getting set up
Use headphones with a microphone. Laptop microphones pick up the room, and the room is what transcription struggles with.
Test it once, alone, on a real question. Two minutes, and it tells you whether your accent and your vocabulary are handled well before you rely on it in a corridor.
Learn the two states. Listening means it is your turn. Speaking means it is not, and talking over it produces a garbled turn that costs a credit and gives you nothing.
Etiquette, briefly
Do not use it in a shared space without headphones. It speaks back, at length, about your business.
Say the whole question. Voice invites trailing off, and a half-finished question gets a half-relevant answer that costs the same as a good one.
Treat the first minute as calibration. How well transcription handles your accent, your colleagues' names and your product vocabulary is something you find out by using it, and it is usually better than expected on the first two and worse on the third.
Did this answer your question?
No, ask a person