Best dictation software for Windows

The fifteen, in one table

We put ours first, and the question it answers is a narrow one: you dictate all day, into whatever window is in front of you, on a Windows PC, and you would rather your voice stayed on it. The other fourteen are listed alphabetically after it rather than scored, because "second best" depends on what you need; the paragraphs below say what each one is for. The table is generated from the same files our comparison pages read, so every figure is whatever its maker last said it was.

Why Deixis is first, and the reasons in plain words

Recognition happens inside the app, on your own processor. Once the speech model is in, the recording, the model and the transcript stay on the machine, on every plan and under every setting, and no mode uploads your audio, so on that point there is nothing to trust us about. The one exception is yours to choose: on the first run you can dictate on Deixis's cloud until the speech model has downloaded.

There is no account. You install it, you press a key, you talk. Nothing local is metered or gated behind a sign-in: all dictation, both speech models, the dictionary and its five vocabulary packs, meeting recording and transcription, screen grabs, screen recording, your own buttons and your own webhooks. Free permanently, not free until we change our minds about the free tier.

One key, everywhere. Hold Ctrl+Win, speak, let go, and clean text lands at the caret of whatever has focus, with your clipboard handed back: your commit message, your email, your ticket. When cleanup calls a language model, only the names of the app and the field go with your words, so each one is punctuated for where it lands; never the window title, never anything on your screen.

The things beyond dictation are the real reason it is first. Double-tap the key for free mode and it keeps listening while you drag a region of the screen or pick a control, then writes one note that interleaves what you said with what you pointed at. Ctrl+Win+S freezes the screen for a grab you can annotate. Ctrl+Win+V records the screen and transcribes what you said in it on your machine; sharing it as a link is the one part that needs a Deixis account. Ctrl+Win+M records a meeting, your microphone and the call both, and transcribes it locally with no key and no network, telling the remote voices apart with a voiceprint that never leaves the machine.

And it is yours to extend. A palette button runs your program with the finished note and types back what it prints; any note, grab or translation can be posted, signed, to a URL you own. Everything it makes is an ordinary file in a folder you own.

Where recognition runs on your own PC

Whisperstream is the closest thing on this page to what we build, and it is good. Its strongest advantage over us is languages: it dictates in 39 of them, on the CPU, where we do seven. It also downloads a small cleanup model into the app, so polishing your prose with a language model needs no key and no server running — with Deixis that is a setup step. If you dictate in more than one language and want it all local, try this one first.

Handy is free and open source under the MIT licence: if you want to read or change the code, pick it. It beats us on languages, with Whisper models covering 99-plus and Parakeet for 25, and on reach: it runs on Linux, Intel Macs and Windows on ARM, where Deixis does not. It is one tool for one job, filling a text box, on purpose. If you dictate in another language or work on Linux, try it first.

Dragon Professional does the thing none of the new tools do: it controls Windows. Custom voice commands and macros, shared word lists, batch transcription of audio files you already have, and a management centre for rolling it out across an organisation. If you need voice to operate the computer rather than fill in a text box, this is a category we do not compete in.

Talon is built for people with limited or no use of their hands, and its own docs say so first. It aims at full desktop proficiency by voice, with every command a script you or the community wrote, and eye tracking for the pointing half. Deixis asks you to hold a key, which is the assumption Talon exists to remove. If holding a key is the problem, Talon is the answer.

The tools that recognise in the cloud

These send your audio out to be recognised, by default or, in Superwhisper's case, whenever you pick one of its cloud voice models. That is a real trade, not only a cost: it buys models we cannot fit on your machine, and it is why each of them does something we cannot. Three of them, Willow, Superwhisper and OpenWhispr, also offer a local option, said so below.

Wispr Flow is the smoothest at following you between devices, desktop and phone alike, and its dictionary can learn from the corrections you make rather than waiting for you to add words by hand. Once the account is made there is very little to set up.

Willow Voice is best at the step after dictation: Scribe turns a rough spoken description into finished prose, which is a different promise from tidying up what you actually said. It also has shortcuts that expand a phrase into a longer saved text, and a shared dictionary for a team on the team plan. Its help centre lists offline dictation on the paid individual plan, without saying which platforms carry it.

Aqua Voice feels the fastest, because Realtime Mode — on its Max plan — shows your words in its toolbar while you are still talking instead of only at the end. It also takes standing custom instructions for how your text should read, and publishes an API for its speech model.

Typeless gives you a tone per application and an "Ask anything" mode on top of plain dictation, with a free plan capped by the week to find out whether that suits you, and it runs on Linux as well as Windows, the Mac and both phones.

OpenWhispr is the one to try if you want the source and the cloud both: it is MIT-licensed, runs on Linux and Intel Macs as well as Windows, and its iPhone app is in the App Store. Setup asks whether to recognise in its cloud or on your own computer, and its pages disagree on which is the default; locally it covers up to sixty languages. It also transcribes files you already have. If reading the code matters or you work on Linux, start here.

Superwhisper is the one built for rolling out to a company: single sign-on, deployment through Jamf or Intune, and control over which models people may use. Its custom modes let you write your own instructions and choose what context each may see, which is more control than we give you; on Windows it runs on ARM64 as well as x64, which we do not. It can transcribe with a local Whisper model, every size of which is on its free plan, and on Pro its AI modes can run on a model on your own PC. One licence covers all its platforms. Its own Windows docs list several features as not there yet.

The built-in one, the Mac ones, and the meeting one

Windows Voice Typing is Win+H and it is already on your PC. Nothing to install or pay, and for occasional dictation it is a reasonable answer, the right first thing to try. Two things Microsoft's own page says about it: it uses online speech recognition and needs an internet connection, and there is no way to add your own words and no custom commands.

MacWhisper and VoiceInk are Mac software, and we list them because half the roundups you will find recommend Mac apps to Windows readers without saying so. MacWhisper takes an audio or video file and hands you a transcript, which Deixis cannot do at all, and it connects to many AI providers under your own keys or to local models through Ollama and LM Studio. VoiceInk's source is published under GPL v3, and its modes switch automatically by application, by website or by a phrase you speak.

Otter.ai is not a dictation tool and does not pretend to be. It is the best answer here to a meeting you cannot attend: its notetaker joins Zoom, Google Meet and Microsoft Teams from your calendar as a visible participant, and everything lands in a searchable shared workspace with sharing into Slack. Deixis records a meeting you are in, on your machine, and nothing joins anything. Different jobs.

Who should not choose Deixis

If you dictate in a language beyond our seven, stop here. Deixis takes English, German, French, Spanish, Italian, Portuguese and Dutch, one at a time (what we measured in each). That is the single biggest gap on this page and several tools above close it.

If you want to hand it a file, it will not take one. Deixis transcribes what it records — your dictation, your meetings, your screen recordings — and there is no drop-a-file transcription in it. Dragon, Superwhisper and OpenWhispr do that on Windows; MacWhisper and VoiceInk do it on a Mac.

If you are on Linux, there is no build and none is planned; Handy, OpenWhispr, Talon and Typeless all run there.

If your PC predates 2013, or runs Windows on ARM, Deixis is not for it. The speech engine is compiled against AVX2, an Intel and AMD instruction set that arrived that year, so an older processor stops with an error rather than running slowly, and there is no build for ARM processors such as Snapdragon.

If you need a particular tool's integration — a calendar link, a bot that attends calls, single sign-on, a management console, a specific cloud provider — buy the tool that has it. And if you want words appearing as you speak, or a tool that reads the text around your cursor, we do neither.