No, not usually — but the reason isn't the one people ask about first. Whether your dictation software sends audio to the cloud matters, but the bigger risk in most real situations is that you just said your password out loud in a room with other ears in it, or a laptop mic that's live on a call. Software privacy and physical privacy are two separate problems, and fixing only one leaves you exposed.
Cloud dictation tools stream your audio to a remote server for transcription. That's a real exposure: your voice data transits a network, sits in someone else's logs, and is subject to whatever retention policy is buried in their terms of service. If you're dictating a password, that string of characters — however briefly — exists on infrastructure you don't control.
But that's the failure mode people fixate on because it's abstract and sounds like a "tech" problem. The failure mode that actually bites people is dumber and more immediate: an open-plan office where the guy two desks over hears you say "my password is Correct Horse Battery Staple, capital C." A shared house with a flatmate walking past. A Zoom call where you forgot your mic was still unmuted from the last meeting. None of that has anything to do with whether your transcription runs on-device or in the cloud — it's just acoustics and who's in earshot.
Both risks are real, but they need different fixes, and conflating them is how people end up feeling falsely safe.
First question: where does the audio go? If you're using a cloud dictation service, assume the audio (and possibly a transcript) leaves your machine and lands on a third-party server, even briefly. That's true regardless of how good their privacy policy sounds. Tools that transcribe entirely on-device don't have this exposure at all, because there's no audio to intercept in transit or store remotely — it never leaves the machine to begin with.
Voxtty works this way: press Alt+D, speak, and faster-whisper transcribes locally using voice activity detection to know when you've started and stopped talking. The words get typed directly into whatever app has focus, via ydotool on Wayland or X11 equivalents. No audio packet ever hits a network socket. If you want optional AI cleanup to strip filler words, that's opt-in and off by default — and even then it's the transcript text going to the Claude API, never audio.
Second question, independent of any software: who can physically hear you right now? This is the one dictation tools can't solve for you. Before dictating anything sensitive — a password, a TOTP recovery code, an API key, a patient's name, a client's financial detail — do a quick check: closed door or open floor plan? Anyone in a video call with an active mic, including your own? Smart speaker or voice assistant in range that might be listening for a wake word? If any answer is uncertain, type it instead. Dictation saves time on prose, not on secrets.
A reasonable middle ground: dictate the surrounding sentence and type the sensitive token by hand. "My new database password is [type it manually] and I've saved it in the vault" gets you most of the speed with none of the acoustic exposure.
Video calls are the sneaky case. Recording indicators lag reality — a call can be recording to cloud storage before the on-screen icon updates, and some clients keep local mic buffers alive briefly after you think you've muted. If you're on a call at all, treat your mic as potentially hot, not off.
Voice assistants are the other blind spot. An always-listening smart speaker in the room isn't recording continuously, but it is listening for its wake word, and misfires happen — a phrase that sounds close enough triggers it and it starts capturing before anyone reacts. If there's a smart speaker or phone assistant within earshot, that's not a room to dictate a password in, on-device transcription or not.
Shared and open workspaces are the everyday version of the same problem: coworkers, flatmates, a video call with participants you don't know well. None of this is solved by choosing local-only software — it's solved by noticing the room you're actually in.
Pick one credential you dictate out loud today — even to yourself when saving a note — and type it manually instead, out of habit. Then check what's actually listening in the room: any open mic on a call, any smart speaker nearby. For the rest of your dictation — emails, notes, code comments, anything that isn't a secret — Try Voxtty free and see what on-device transcription looks like when the audio genuinely never leaves your machine.
Local-first voice dictation for Linux. Press Alt+D, speak, and your words land in whatever app has focus — nothing leaves your machine.
Try Voxtty free →