You press the hotkey, you speak, the tool says it transcribed you — and the remote window stays empty. This is not a bug in your microphone. It is a consequence of how remote sessions carry keystrokes.
Windows offers more than one way to synthesise a keystroke. Most dictation software
uses SendInput with the KEYEVENTF_UNICODE flag, which says
in effect "a character arrived" and carries the character itself. It does
not correspond to any physical key, so it has no scan code — the
hardware-level number a real keyboard emits.
That works everywhere locally. Remote desktop is different. Citrix, RDP and VMware Horizon do not forward characters; they forward scan codes, because the remote machine needs to apply its own keyboard layout. When a synthetic Unicode event arrives at the client, there is no scan code to send onward.
The remote application receives nothing at all. No error, no partial text — silence. Which is why this looks like the dictation tool is broken rather than incompatible.
The same failure shows up in some Java and Qt applications, which read scan codes directly, and in anything running a low-level keyboard hook that filters injected input.
Put the text on the clipboard and send Ctrl+V. The remote application performs its own native paste, so it works wherever paste works.
The catch: inside a remote session, Ctrl+V pastes from the remote clipboard, which only matches your local one when clipboard redirection is enabled. Locked-down legal and healthcare environments routinely disable it for data-loss prevention — precisely the places most likely to hit this problem in the first place. It also briefly takes over your clipboard.
Look each character up against the target window's keyboard layout, find the virtual key and shift state, convert to a scan code, and send that. The remote session forwards it like any real keypress. No clipboard involved, so policy cannot switch it off.
The catch: a scan code only exists for characters reachable on the active layout. An em dash, a curly quote, an accented letter on a US layout — none of those have a key to press, and have to fall back to Unicode individually, which the remote session will still drop.
Sidesteps the problem entirely, and is often impossible: the remote environment is usually locked down, has no microphone passthrough, or forbids installing anything.
DictaFlow handles Citrix and VDI dictation too, and has done for longer. If you want a subscription, cross-platform support and a vendor with this as their whole pitch, it is a reasonable thing to buy and you should look at it.
The difference worth knowing is where your audio goes. DictaFlow processes speech in the cloud, on a monthly subscription. Recount processes it on your own machine and is bought once.
That distinction matters most to exactly the people who end up dictating inside Citrix in the first place: if the work is sensitive enough to live on someone else's infrastructure, sending the audio to a third party may be the thing your IT department objects to.
| DictaFlow | Recount | |
|---|---|---|
| Price | $7/month | $29.99 once |
| Three years | ~$252 | $29.99 |
| Speech processed | In the cloud | On your machine |
| Works with no internet | No | Yes |
| Platforms | Windows, Mac, iOS, Android | Windows only |
| Citrix / VDI | Yes, and longer | Yes, untested in a live session |
If you want it on a phone or a Mac, or you would rather someone else carried the maintenance, pay the subscription. The case for buying once is narrower than it sounds: it is for people who will still be dictating in three years, on Windows, and would prefer the audio never left the building.
Recount ships all three insertion methods and lets you switch between them from the tray, because no single one reaches everywhere:
If dictation stops working in one particular application, you change one setting rather than changing product.
Stated plainly: the scan-code path is built and its layout lookup is under test, but it has not yet been verified inside a real Citrix session. If you try it there, the fourteen-day trial costs nothing and I would genuinely like to know the result either way.
People dictating inside Citrix are usually doing so because the work is confidential enough to live on someone else's infrastructure. Recount runs the speech model on your own machine — there is no upload path in the software, so there is no cloud vendor to add to a risk register and no business associate agreement to negotiate.