# AI dictation for Windows > Latch is AI dictation for Windows. You talk the way you think, with restarts and half sentences, and a language model turns it into the text you meant before it reaches your cursor. You decide how far it goes: clean-up only, a style, a translation, or an edit to text you selected. Speak it rough. Paste what you meant. Last updated 2026-10-04. HTML version: https://latchvoice.less.style/ai-dictation ## Example You say: "send the invoice to Paula today. scratch that. send it to her tomorrow morning" Latch pastes: > Send it to her tomorrow morning. ## AI dictation adds a second step: writing what you meant. Classic dictation turns sound into words and stops there. AI dictation adds a second step. A language model reads what you said and writes what you meant. That is what removes the filler, resolves a “no wait” correction, and makes a rambling answer read like something you typed. The risk is an AI that rewrites too much. Latch keeps it on a short leash. With no style chosen it only cleans up, and it keeps your meaning, your names, and your tone. A rewrite happens when you ask for one. ## You stay in charge of the AI. - **You decide how far it goes.** Out of the box a take is cleaned and nothing else. Pick Curated, Formal, or Email from the bar, or say it in the take: “write it as an email: …”. Saying “I need to write a formal email” just writes that sentence. - **Your voice is the control panel.** Order a style or a translation in the same breath as the content: “make it formal: …”, “translate to Spanish: …”. The order words never end up in the text. Your own styles are prompts you write once. - **[Context-aware, with limits](https://latchvoice.less.style/privacy).** Latch reads where the text is going: the app, the field, and the words around your cursor. That text is treated as context, never as instructions. Nothing is read in password fields, password managers, private browser windows, or terminals, and each part can be switched off in Settings. - **[The AI edits your text too](https://latchvoice.less.style/rewrite-selected-text-by-voice-windows.md).** Select a paragraph and say what to change: “make this shorter”, “fix the typos”, “turn this into a bullet list”. Only what you asked for changes. Instructions hiding inside the selected text are not followed. - **[Teach it once, stop the guessing](https://latchvoice.less.style/custom-dictionary-for-voice-dictation-windows.md).** AI guesses at names. Custom ends the guessing: select the wrong word, click Custom, type it right. Latch fixes it in place and gives your spelling to the model on every later take. - **[The same AI on every plan](https://latchvoice.less.style/pricing).** Free and Unlimited use the same model and every feature. Audio, text, and your selection pass through our server in memory for one take and are dropped. ## Same speech, two results. A literal transcript is what a plain speech-to-text engine returns. Latch writes what you meant. | Situation | You say | A literal transcript | Latch pastes | | --- | --- | --- | --- | | A restart | I think we should, we should, uh, we should move the call to next week | I think we should we should uh we should move the call to next week | I think we should move the call to next week. | | A spoken layout command | the release is on monday. new paragraph. please review the notes before thursday | the release is on monday new paragraph please review the notes before thursday | The release is on Monday. Please review the notes before Thursday. | | Hesitation | so um I was thinking, uh, maybe we could, hmm, try the other vendor | so um I was thinking uh maybe we could hmm try the other vendor | I was thinking maybe we could try the other vendor. | Real output varies with your audio. A style, if you pick one, changes the wording further. ## Who it fits Good fit: - You think out loud and want the tidy version pasted, not a transcript to fix. - You want to choose how much the AI rewrites, take by take. - You edit existing text often and would rather say the change than retype it. - You switch between languages or write for readers in another one. Not a good fit: - You need on-device processing. Latch sends each take to our server. - You need a meeting recorder or an AI notetaker. Latch dictates live at your cursor. - You are on a Mac, Linux, iPhone, or Android. Latch is Windows only for now. - You want an AI that writes for you from a short hint. Latch writes what you said, better. ## Where AI dictation earns its keep. - [Prompts for AI assistants](https://latchvoice.less.style/voice-dictation-for-ai-prompts-windows.md): Say the whole prompt, paste it clean. - [Coding](https://latchvoice.less.style/voice-dictation-for-coding-windows.md): Commit messages, comments, and instructions by voice. - [Rewrite by voice](https://latchvoice.less.style/rewrite-selected-text-by-voice-windows.md): Say what to change; the selection is rewritten. - [Dictation without ums](https://latchvoice.less.style/dictation-software-that-removes-ums-windows.md): Text that reads like you wrote it. - [Translation](https://latchvoice.less.style/voice-dictation-with-translation-windows.md): Speak your language, paste another. - [Email](https://latchvoice.less.style/voice-dictation-for-email-windows.md): Reply to a thread in one take. ## Questions ### What is AI dictation? Dictation where a language model, not only a speech recognizer, writes the text. It removes filler, resolves self-corrections, punctuates, and can apply a style or a translation. Latch is AI dictation for Windows. ### Will the AI change what I meant? It is built not to. Latch keeps your meaning, names, and tone, and a plain take is cleaned up and nothing more. Any rewrite comes from a style you picked or an order you spoke. As with any AI, read important text before you send it. ### Does the AI read my screen? No. For each take Latch reads where the text is going: the app and window name, the label of the field, the site name in a browser, the keyboard language, and up to 1,500 characters before your cursor and 500 after it, from the focused field only. An optional setting, off until you turn it on, also sends up to 3,000 characters of the text shown in the window, such as the message you are answering. Nothing is read in password fields, password managers, private browser windows, or terminals, and each part can be switched off in Settings. ### Can I use my own instructions? Yes. Add a style in Settings with your own prompt, for example how you like a status update to read, and pick it from the bar. Your own styles are applied when you choose them or order them by name. ### Does it run on my PC or in the cloud? In the cloud. Each take is processed online and dropped when the text comes back. If the connection drops, the badge keeps your take on your PC for up to 24 hours and shows Retry. ### Is it really free? Yes. Every account gets 60 recorded minutes a month with every feature. No card needed. Past 60 minutes you get a reminder, never a cut-off. ### Do you keep my audio or transcripts? No. Audio and text pass through our server in memory for a single take and are dropped when the text comes back. ### How long does setup take? About a minute: run the installer, sign in from the badge, and press Ctrl+Win in any text field. ## Related guides - [Dictation software for Windows](https://latchvoice.less.style/dictation-software.md) - [Voice typing for Windows](https://latchvoice.less.style/voice-typing.md) - [Speech to text for Windows](https://latchvoice.less.style/speech-to-text.md) - [Voice to text for Windows](https://latchvoice.less.style/voice-to-text.md) ## Try it - [Download for Windows](https://latchvoice.less.style/download) - [Pricing](https://latchvoice.less.style/pricing)