The pre-meeting check

Preflight answers one question: if you started this meeting right now, would it work? Not "is Whispr installed correctly" — that is whispr doctor — but is this microphone loud enough, is this meeting's material loaded, will these accounts still be answering in twenty minutes, and what will it cost.

Open it from the Preflight tab, from Check before starting on the Live view, or from Check on a meeting card. It takes about a minute, and most of that is reading a sentence out loud.

Nothing on this screen can stop a meeting starting. The Start button sits at the top of it, beside whatever warnings there are, and it is never disabled. A meeting happens once, and being refused at 09:59 is worse than a generic answer.

Quick answers

Some questions go to one cheap engine instead of all of them — by default the ones about logistics and background, which is the saving that setting exists for. Preflight names which engine that is.

On a sales call it warns, because that same lane carries pricing, contract, procurement and timeline — the questions a buyer most needs answered precisely. Whispr does not change your engine choice for you; it tells you before the call so you can, in Settings.

On Windows, run this from the app

whispr preflight in a terminal cannot measure your microphone on Windows, and says so rather than guessing: Whispr captures audio inside its own window there, so a command line has nothing to open one with. Everything else on the page — the meeting, the credentials, the headroom, the costs — is checked exactly the same way.

Use the pre-meeting screen in the app instead. It runs the identical measurement through the identical capture.

Your equipment

Microphone says which input Whispr will record you from. Three answers are worth reading twice:

• Following the system default. A legitimate choice, and usually the right one — but "the default" is whatever the operating system is pointing at, and a headphone jack with nothing plugged into it is a perfectly valid default that records silence. If you have more than one input, pick the one you mean.
• Not connected. You chose a headset that is not plugged in. Recording falls back to the default and your choice is kept for when it comes back.
• Chosen but something else is recording. A meeting is running and started on a different input. Stop and restart it to switch.

Microphone level is the check that does not exist anywhere else, and the one most likely to save a meeting. Press Measure it, read the sentence out loud for six seconds, and it reports one of:

• too quiet to transcribe — speech at this level sits in the microphone's own noise. This is not "slightly worse accuracy"; it is no transcript at all, and therefore no questions detected and no answers. Raise the input level — on a Mac Whispr can do it for you, or System Settings › Sound › Input; on Windows it is Settings › System › Sound, and Whispr cannot set it for you there, because nothing it can reach touches the operating system's input volume. Drag Input volume up and watch the meter move as you speak. If it is already at maximum, use a different input.
• quiet but workable — recognised, with no margin. Fine in a quiet room, loses words in a noisy one.
• good — nothing to do.

It is a different measurement from whispr verify audio, deliberately. That command asks whether audio is arriving, which is the right question for a dead microphone or a permission problem. It passes on an input that is wired and far too quiet — measured on a real Mac: speech peaking at 254 out of 32767, a green verify audio, and a spoken mock interview that transcribed nothing at all. Preflight asks the stricter question.

Hearing the other side cannot be proven from this screen, because nothing is playing. It is proven the moment a real meeting works, or by running whispr verify audio with something playing through your speakers.

Proven to work

One row per credential, and the wording matters:

• proven — this exact configuration has really transcribed, or really answered, and recently.
• saved but never tested — the key is there and nothing has ever used it. Neither "ready" nor "missing"; press Test it.
• not configured — there is no credential. Add it in Settings.

A proof is tied to the configuration it was taken against, so changing a key or a model retires it and asks you to test again. A row can also say a check is running.

What this meeting is grounded in

The material this kind of meeting wants — a CV and the job advert for an interview, what you sell and who you are selling to for a sales call. Missing material never stops a meeting; it changes the answers, from specific to generic. The one exception is a sales call with nothing loaded about what you sell, because nobody in that room can correct an invented product.

When the other side arrives on a cable

If you have set the other side to come from an audio input, two extra rows appear and one changes.

Hearing the other side names the device instead of talking about screen permissions, and — unlike the normal setting — it can actually be proved before the call, because there is somebody on the other end who can make a noise. If the device is unplugged it fails plainly, and the fix is the cable, not a permission.

Their level is the row worth reading. A cable in the wrong socket, or an interface gain at zero, produces a channel that is wired, passes every other check, and cannot be transcribed — the same failure your own microphone can have, one channel along. Raise it on the interface; Whispr cannot.

And if one device is set as both your microphone and where the other side arrives, that is called out as a failure: both channels would hear the same audio and each side would be recorded as the other.

Answer pace

If your answer engines have been taking too long to start, this says so — per engine, with the median time to their first word. It is a warning, never a refusal, and it is read from what those engines have already done rather than timed here, so it costs nothing and spends nothing.

The number to know: there are about three seconds between the other person finishing and you being expected to speak. An engine past that wrote something worth reading afterwards and nothing you could use in the moment. See Settings → answer engines for what to do about it.

How much is left

Asked only when you press a button — these are billing endpoints on your own accounts, so nothing polls them.

Most providers publish no remaining quota at all, and the row says exactly that rather than showing a tick. One case is worth knowing: on Cloudflare's free plan, Workers AI refuses inference past the daily allowance rather than billing it, so answers stop mid-call with everything else still green.

Nothing here refuses a session either. It is a warning early enough to act on.

What it will cost

One line per provider account, in the unit it is billed in, with what a 45-minute call typically consumes. Two engines on the same account share one bill, so they share one row. If a switched-on option has no credential stored, the row says so — that engine would simply be skipped.

From the terminal

whispr preflight prints the same verdicts, asks for the same sentence, and exits non-zero if something would stop the meeting working. It reads the stored proofs rather than re-running them; add --check to prove the credentials for real, or --no-speech-test to skip the microphone measurement.

whispr preflight               # the whole check, with the guided measurement
whispr preflight --check       # re-prove the credentials too
whispr audio-devices           # what this machine offers, and what is chosen (macOS only)
whispr audio-devices --set <id>   # choose one from the terminal

What the level check also does

Reading the sentence out loud measures whether your microphone is loud enough to be transcribed — and it does a second thing worth knowing about.

It is the only moment Whispr learns how loud you are, on this microphone, at this gain. That reference is what lets it tell your voice from the room: a television, a colleague at the next desk, a call on another device. All of those reach your microphone, and your microphone is structurally you — so without a reference they are recorded as things you said. Measured in a real meeting: thirty-eight turns attributed to the user while they said nothing, including a news broadcast about the CIA and a report from Florida.

The test is distance, not loudness, and that distinction is the reason it works. Speaking into your own microphone puts you about 20 cm from it; a television is three metres away, which is roughly 25 dB quieter — measured here at 28. So a turn far below your own measured level is the room, whatever its absolute volume. There is no absolute number that would work: on one Mac real speech measured 254 while ambient noise on another setup measured 478.

Two consequences. If you never run the level check, none of this happens — nothing is dropped, and the app behaves exactly as it did before. And when turns are left out, Whispr says so and names the cause, because "audio was dropped" sends people to change their input device, which is the one thing that will not help.