Hold a key, talk, let go. Your words, cleaned up, in your style. With a mark that shows it was spoken, not generated.
Chrome extension, beta. YouTube, LinkedIn, X, Reddit, Gmail, every text field.
#dictapost the spoken mark, on for public sitesStored: a 512-byte fingerprint, never the text. More in the questions.
Five posts on a site and it knows your register there. Examples stay in your browser.
Two modifier keys. Never fires while typing.
Typed at your cursor, or shown first. A thumbs down teaches the next version.
Typed text works too: select it, press Ctrl+Shift+E. Any language the models know.
Prices include VAT where it applies. Cancel any time. Buy here or inside the extension, same account.
Leave your email and you get one message when Dictapost is in the Chrome Web Store. No newsletter.
Tidy does not: spelling, punctuation, fillers, sentence breaks, wording and order stay. Shape turns a spoken thought into a post: it may reorder, cut repetition and form paragraphs, but every claim, every opinion and the wording come from what you said. Neither mode adds a hook, a closing line or nicer vocabulary. If one does, the thumbs-down button is how you report it.
They polish. Their own words: "polished text", "auto-edits speech". Dictapost shapes only from what you said and checks afterwards how many of the words are yours. Their style memory is a black box; Dictapost's examples are texts you can read and delete. The server keeps nothing. And you can run it on your own model.
The spoken mark. The post was dictated through Dictapost, and the id points to a record: when, how many seconds of speech, how many words. Open dictapost.com/v/ followed by the id, paste the post, and you see how much of it came out of that dictation. With the extension installed, the number shows next to the mark in YouTube, LinkedIn and X feeds. The mark is on for LinkedIn, short posts, YouTube and forums, off for email and chat. Switch it per site in the options, and delete the line from any post before you send.
Time. Two hundred words take about ninety seconds to say out loud; a generator takes two. The record holds the seconds of speech, the time, the word counts and a 512-byte fingerprint of the words, signed and append-only. The text cannot be read back from it; the audio is never kept. The number on the badge says how much of the posted text came out of that dictation: cut a line and it stays at 100, paste in a generated paragraph and it drops.
Copied from another post: the check shows 0 %, because the fingerprint belongs to other words. Made up: an id Dictapost never issued shows as unknown. Read aloud: yes, someone can read a generated text into the microphone. The mark does not prove a human thought the words. It proves someone spent the seconds saying them, and the record shows those seconds.
Any language the models handle, which covers all major European languages and many more. Output is in the language you spoke. Mixed languages stay mixed.
Not yet. Dictapost is a Chrome extension. Firefox follows. A desktop companion for Mac and Windows is on the list.
Pro at 6 euros a month. Founders keep 12 euros a year as long as their subscription runs.
One person, who measured how he speaks and how he writes and found they are different languages. Spoken sentences ran 15 words, written ones 9. Dictation alone never sounded like him. So he built the part in between.
I tested the new model on the same 40 questions and it is actually worse than the old one, 3 points worse, which is not what I expected. So: don't upgrade yet.