Dictation in QUILL, QUILL Lite and QUILL Transcribe

Press Ctrl+F11, talk, and pause. Each time you pause, what you said is written into the document at the cursor, a soft tone tells you it arrived, and the words are read back so you know they are right. Then it keeps listening. The speech recognition ships inside the programs: there is nothing to download, no account to make, nothing you say leaves your computer, and no recording is kept anywhere. Bigger models and cloud services are there if you want them, and only if you choose them.

This page covers all of it: every key, every phrase dictation understands, every setting, every speech engine and model and how to choose between them, and what to do when it mishears you. It covers QUILL and QUILL Lite, and tells you wherever they are different. QUILL Transcribe, which dictates into other programs, uses the same engines, models, words and commands; its own guide covers its keys.

Your first dictated sentence

This works the same in QUILL and in QUILL Lite. Nothing needs setting up first, and you do not need to say any punctuation. Before you start, the way to stop is Ctrl+F11 again, or saying "stop dictation", or Escape to throw away the phrase being heard. It helps to know that first if you cannot see whether the microphone is on.

1. Put the cursor where you want the words. If you select some text first, the first phrase you say replaces it.
2. Press Ctrl+F11. The first time after opening the program it takes a second or two while the speech engine loads. Then you hear two rising tones and "Dictation on".
3. Say a sentence the way you would say it to a person, and pause.
4. A moment later the sentence is in the document, you hear one short soft tone, and the words are read back to you.
5. Keep going. There is nothing to press between sentences.
6. When you are done, press Ctrl+F11 again, or say "stop dictation". You hear two falling tones and "Dictation off".

A phrase is written once you have been quiet for a little under a second, so a breath in the middle of a sentence does not cut it in two. If a pause does split a sentence and the next part starts with a word like and, but, which or to, the full stop the pause put in is taken back out and the sentence carries on.

Every phrase is one step for Ctrl+Z, punctuation and capitals included, even a phrase that replaced a selection.

Holding Ctrl+F11 to talk

Press Ctrl+F11 to start dictating and press it again to stop. That is all most people ever need, and it is how dictation starts out.

If you prefer to hold the key while you talk, like the button on a walkie-talkie, turn on Hold the dictation key to talk; a quick press still turns it on and off in More Dictation Settings. Then:

1. Put the cursor where you want the words.
2. Press and hold Ctrl+F11. You hear the two rising tones and "Dictation on" straight away.
3. Keep holding the keys, and say what you want to say.
4. Let go. Dictation writes your last phrase, then you hear the two falling tones and "Dictation off".

With that turned on, a quick press still works just the same: dictation turns on and stays on until you press Ctrl+F11 again. Half a second or more is a hold. Stopping never cuts you off: if you press or let go while you are still finishing a sentence, dictation writes that phrase first.

What you hear

Four moments, each with its own tone and its own words. Both halves can be changed, and two of them can never be silenced.

Every row can be changed in Dictation Settings, and every tone can be changed or silenced in Tools, Sound Scheme. Two things are never silenced: a failure is always spoken, and so is the answer to a spoken command, because both answer a question.

The soft tone plays only once the words are really in the document, so hearing it means they arrived. If you have the words read back, they follow a quarter of a second after they are written, so your screen reader's own reaction to the new text does not cut them off.

Use headphones if the read-back is on. Through speakers, the microphone can hear the read-back and write it down a second time. With speakers, set "After each phrase is written, give me" to a sound only.

How to tell it is listening, without sound and without sight

Every outcome above is also written down, so none of this depends on hearing a tone. Two ways, and both reach a braille display. The status bar has a Dictation part that says Off, Listening, Hearing you, Writing, Spelling or Waiting for wake phrase, and F6 moves focus there so you can read it cell by cell. And Tools, Dictation, Dictation On is a check box in the menu, so opening that menu announces it as checked or unchecked.

The same holds for failures: a failure is always spoken and written to the status bar, whatever the sound settings say, and it is never reduced to a tone. If you would rather have no tones at all, silence them in Tools, Sound Scheme and nothing is lost.

Hearing the punctuation in the read-back

When a phrase is read back to you, its punctuation is read too, using the same names you would say to dictate it: "Hello comma world period". Quotes, question marks and new paragraphs are said the same way: "open quote", "question mark", "new paragraph". You hear them whatever your screen reader's punctuation level is, so you know exactly what was written without stopping to check.

A few things stay quiet, because saying them would only get in the way: an apostrophe or a hyphen inside a word, and the full stop or comma inside a number such as 3.5 or 1,000. A web address is said the way you would dictate it, so example.com is read as "example dot com". In Spanish you hear the Spanish names, such as "coma" and "punto".

The status bar and a braille display show the real characters, as in "Dictated: Hello, world." If you would rather hear just the words, uncheck Say punctuation marks in the read-back in More Dictation Settings. It is checked to begin with.

The two kinds of dictation

QUILL has two, and they work quite differently. QUILL Lite has the first one.

Live Dictation: writes as you go

Ctrl+F11, in both products. Each phrase is written the moment you pause, so you hear it land and can correct it by voice straight away. This is the one described above and for the rest of this page unless it says otherwise. In QUILL it is under Tools, Speech, Live Dictation; in QUILL Lite it is Tools, Dictation.

Locked Dictation: records a passage, then writes it out

Ctrl+F9, QUILL only. Press it once and speak as long as you like, across sentences and paragraphs; press it again and QUILL transcribes the whole passage and inserts it as one undoable edit. It uses QUILL's offline speech engines on this computer, so nothing is uploaded here either, and it needs no speech model to start (which engine Locked Dictation uses). Use it when you would rather talk without interruption and read the result afterwards.

• While recording, Escape stops and keeps what you said for transcription; Shift+Escape cancels and discards it.
• A session stops by itself after five minutes, and stops and preserves your audio if QUILL loses focus.
• Audio is saved to a recovery folder before transcription starts, so a crash cannot lose a passage. Tools, Speech, Locked Dictation, Dictation History and Review lists every recording whose transcript was never inserted, so you can insert, copy or discard it. If anything is waiting at startup, QUILL says so.

Dictate (Offline): a simple on and off switch

QUILL only, in Tools, Speech, on Ctrl+Shift+Grave, Shift+D. Start it, talk, stop it: QUILL transcribes in the background and inserts the text at the cursor as one edit, with the word count in the status bar. It shares its engine and microphone with Locked Dictation.

Live Dictation and Locked Dictation each remember their own microphone choice.

Every dictation key

These are the keys it starts with, and you can change any of them in QUILL Lite's Keyboard Manager (Ctrl+Alt+Shift+R) or QUILL's Settings, Keyboard. Each row starts with what you want to do, then the key to press.

Keys both products share

These keys are the same in QUILL and QUILL Lite, so a key you learn in one works in the other.

Keys only QUILL has

Locked Dictation and the offline switch. QUILL Lite keeps things simple and has Live Dictation only.

Ctrl+Shift+Grave, Shift+D is two presses, not five keys at once: press the QUILL Key (Ctrl+Shift+Grave), let it go, then press Shift+D. QUILL Lite does not use two-step keys like this. To see what else is different, read how QUILL and QUILL Lite compare.

The speech engines

Choose the engine in Dictation Settings (Alt+Shift+F6), in the Speech engine list. The same list is in QUILL, QUILL Lite and QUILL Transcribe. Moonshine and Whisper ship inside all three, so they work with no download and no connection. Models you download, and cloud services you set up with your own account, join the list. Help with picking one is in Choosing an engine.

Every engine except Windows speech recognition puts in its own punctuation and capitals, so with them you just talk.

Which engine Locked Dictation uses, in QUILL

QUILL's Locked Dictation (Ctrl+F9), Dictate (Offline), Transcribe Audio or Video (Offline), Generate Captions (Offline) and watch folders share a second set of engines, also on this computer. You manage them in Tools, Speech, Speech and Dictation... (Alt+T, H, S), which opens Manage Speech Models on its "Dictation (Offline)" tab. To be walked through it instead, choose Help, Download Optional Components... (Alt+H, D) and the row Dictation (offline speech). There are three engines, in this order:

1. Whisper.cpp, the recommended default: light and fast, and works on any computer. The engine itself is small, about 8 MB, and downloads from QUILL's own checked release the first time it is needed. Its models then download in Manage Speech Models, from Tiny (75 MB) and Base (145 MB) to Small (465 MB), the suggested starting point, and larger.
2. QUILL speech models (Parakeet, Nemotron and more): the very same models as Live Dictation, QUILL Lite and QUILL Transcribe, in the same folder, so a model you already downloaded is never downloaded twice. This one choice replaced the separate Parakeet and Nemotron engines QUILL used to list.
3. Faster Whisper: the most accurate, larger, and the only one that can use a graphics card. It installs when you ask for it, where this copy of QUILL can install it.

Which one QUILL uses: an engine you chose yourself always wins. Otherwise an installed Parakeet is used, because when you are silent Whisper sometimes makes up words, like a stray "thank you", and Parakeet only writes what it really hears. Otherwise whisper.cpp, once it has a model. And until there is any model, the built-in Moonshine is used, so Locked Dictation needs no model: with whisper.cpp as the engine, the first press of Ctrl+F9 only offers the small whisper.cpp engine itself. Choose QUILL speech models as the engine and nothing is downloaded at all; Dictate (Offline), transcription and captions then work with nothing downloaded too, because the built-in Moonshine and Whisper count as that engine's models. For the same recording and the same model, Locked Dictation and Transcribe a Recording write the same text. If a recording has no speech in it at all, QUILL tells you so instead of guessing.

Voice Command, Voice Conversation Mode and Listen for Hey QUILL have their own Voice recognition engine setting: follow the main engine, whisper.cpp (accurate), or Moonshine tiny (fast, light). Vosk is no longer a dictation or voice engine; QUILL keeps it only for finding chapters in podcasts.

How the default was chosen

The full, current evidence is on its own page: why dictation recommends what it does, with every model measured on real people's speech in October 2026, how we tested, how we reasoned, and why your own results will differ. The table below is the first, earlier comparison that chose the built-in default, kept for the record; its figures came from a synthetic voice, which flatters every engine.

We tried eight engines on ten everyday sentences each, on 25 September 2026. We wanted one that keeps up with you on an older Windows 10 computer, puts in its own punctuation, and is small enough to come inside the program.

Anything under 1.0 in the speed column keeps up with you as you talk. Moonshine tiny was the most accurate and by far the fastest, so it works well even on an older computer. Whisper tiny is there for a voice or microphone that Moonshine struggles with. Parakeet 3 is very good but five times the size, so it is an optional download in Speech Models rather than something packed into every installer. Vosk was dropped from dictation on 13 October 2026: it does not punctuate, knows only English, and has had no new release since 2022.

One thing to keep in mind. For this test, Windows' own computer voice read the sentences. That is easier to understand than a real person, so every engine will make more mistakes with your voice than the table shows. Windows speech recognition in particular does much worse with a real voice. Moonshine tiny is still the default; the advice about which model to download now comes from tests with real people's recordings, in Tested on real speech.

Choosing an engine

The short answer: start with Moonshine, the one dictation starts with. Change only if it often mishears you, and then try the engine that matches what matters most to you.

Every choice is a trade. A bigger model makes fewer mistakes with accents, names, fast speech and noisy rooms, but it needs more memory, takes longer to load, and keeps your text waiting a little longer after you pause. A cloud service can be very accurate on any computer, but your speech leaves it. This table puts each concern next to the engines that answer it.

What the tier labels mean

Every model in the Speech engine list and in Speech Models carries one of three tiers, for example "Moonshine (fast, punctuates for you), Lightweight". The tier says how hard the model works your computer, not how good it is.

• Lightweight: keeps up comfortably even on a modest computer. The built-in Moonshine and Whisper, Moonshine base, the two Parakeet models (NVIDIA Parakeet TDT 0.6B v3 and NVIDIA Parakeet Unified 0.6B) and Whisper base.en.
• Balanced: more work than the built-in engines; keeps up on a computer with four or more cores. Nemotron and Whisper base.
• High accuracy: the most work; needs a capable computer, or patience. Whisper small and larger.

The tiers follow each model's real speed on real recordings, measured on 22 October 2026, so a tier can be Lightweight and still among the most accurate: the two Parakeet models are both.

A recommendation for this computer

You do not have to work this out alone. Dictation Settings shows a read-only line, Recommended for this computer, with the model that suits your memory and processor and the reason. In Speech Models you can say what matters most and get a recommendation for that:

1. Open Dictation Settings (Alt+Shift+F6) and press Better Accuracy: Speech Models... (Alt+B).
2. Move to What matters most (Alt+W) and choose a balance, the quickest response, or the fewest mistakes.
3. Read the Recommendation box (Alt+C). It names a model and says why.
4. Press Show Recommended Model (Alt+S) to move to that model in the list, where Download and Use for Dictation are.
5. Once it is downloaded, Switch to Recommended Model (Alt+H) chooses it and changes nothing else. Press OK in Dictation Settings to keep it. If it is not on this computer yet, nothing is downloaded: you are told, and taken to its row.

In QUILL Transcribe the same advice is in Model, Choose Model... (Ctrl+M), as Recommended for this computer (Alt+C), and in its Set Up window. A recommendation is only advice: it never changes your engine. Its figures come from our tests with real people's recordings (Tested on real speech), on one computer, so your own voice and computer may still differ. OpenAI and the speech providers are never recommended, because they send your speech away.

It also counts memory. On a computer with 8 GB, dictation may use up to 1.6 GB beside your screen reader, and never more than is free when you ask, less half a gigabyte. So on a busy 8 GB computer the advice can name a lighter model than the table below.

Switching engine, and a worked example

To switch, open Dictation Settings (Alt+Shift+F6), move to Speech engine, arrow to the one you want, and press Enter for OK. The next time you press Ctrl+F11 it uses the new engine, and Alt+F9 tells you which one is in use. Choosing OpenAI or a speech provider first asks whether your speech may be sent; No is the answer if you just press Enter, and your engine stays as it was.

For example, Sam writes reports on an 8 GB laptop with four cores, and Moonshine keeps getting the names of Sam's colleagues wrong.

1. Sam first adds the names to My Words and Phrases (Alt+Shift+F10). That fixes most of them on any engine.
2. For the rest, Sam opens Speech Models (Alt+Shift+F6, then Alt+B), sets What matters most to the fewest mistakes, and reads the recommendation: NVIDIA Parakeet Unified 0.6B.
3. Show Recommended Model moves to Parakeet Unified. Sam presses Download..., reads the question naming its 663 MB size and licence, and answers Yes.
4. When it is done, Sam presses Use for Dictation, then OK in Dictation Settings, and dictates as before. Each phrase is written when Sam pauses, as with Moonshine.
5. On battery, when the laptop is too busy to keep up, dictation switches to the built-in engine for a while and says so; Sam's choice stays Parakeet Unified.

Better accuracy: optional speech models

Dictation works the moment you install. If you want more accuracy, you can download a larger model: the same local models VS Code offers, plus the rest of the Whisper family. Free, optional, and run on your computer's processor; no graphics card needed.

Your voice never leaves the computer. The only thing that comes over the internet is the model itself, once, when you ask for it. Nothing downloads on its own, and the built-in Moonshine stays your engine until you choose another. QUILL, QUILL Lite, QUILL Transcribe and Quill Inkwell all have this window, in the same place in Dictation Settings, and share the models on one computer, so a model is downloaded once. QUILL Transcribe also opens it from Model, Model Library... (Ctrl+L). Every model runs on the processor; a graphics card is not needed and changes nothing.

Downloading one

1. Open Dictation Settings (Alt+Shift+F6) and press Better Accuracy: Speech Models... (Alt+B).
2. Arrow through the list. Each row says what the model is good for, its size, and whether it is here. Details says what it is better at, its languages, download and disk size, which computers suit it, whether this computer should keep up, its published accuracy, its licence and where it will be saved.
3. Press Download.... A question names where it comes from, its size, its licence and the folder. Answer Yes. On a metered connection you are asked about that first, and if the drive is too full you are told before anything starts.
4. You hear "Downloading", then every quarter of the way. Cancel Download stops it and keeps what arrived, so Download carries on later. Every file is checked against a recorded checksum before it is used.
5. Press Use for Dictation, then OK in Dictation Settings. Or choose it in the Speech engine list, where downloaded models show "(downloaded)".

Where they go: %LOCALAPPDATA%\QuillVille\Dictation\models, one folder for every program. In a portable copy, models are saved inside the portable folder, so everything travels together. Remove in the same window frees the space and says how much. When you switch away from a downloaded model and press OK, a question offers to delete the old one to free its space, and says that QUILL, QUILL Lite and QUILL Transcribe share its files; No is the answer if you just press Enter. It is not asked while dictation is running. If a downloaded model goes missing, dictation says so in one sentence and carries on with the built-in engine; it never switches silently.

Will my computer keep up? When the window opens it spends about two seconds timing the built-in Moonshine, then tells you for each model whether this computer should keep up comfortably, keep up but leave your screen reader less room, or "may lag behind your speech on this computer". You can still choose any of them.

Words appear when you pause with every downloaded model except Nemotron. Nemotron recognises while you speak, so the words heard so far show in the status bar and on braille as you talk (Seeing the words while you speak), and each phrase is written when you pause.

The models, in our suggested order

The order is our suggestion: the models that did best on real speech first, then the rest of Whisper, smallest first.

Five English-only models were removed in this release: Whisper small.en, Distil-Whisper small.en, Distil-Whisper medium.en, Whisper medium.en and Distil-Whisper large-v3. On real recordings, Parakeet Unified got 3.2% of words wrong on the held-out audiobooks, against 3.6% to 6.3% for these five, and ran faster on one thread, in the same or a smaller memory tier. Whisper base.en and every many-language Whisper stay.

If you were using one, dictation moves when you update: to Parakeet Unified if it is already downloaded, otherwise to the built-in model. Nothing is downloaded. The next time dictation starts it says one sentence, once across every QuillVille program, for example "Whisper small.en was removed in this release. Dictation now uses Parakeet Unified, which you already have." or "Whisper small.en was removed in this release, so dictation uses the built-in model. Parakeet Unified, in the model list, is more accurate." To choose Parakeet Unified, find it in Speech Models, then Download and Use for Dictation.

Your downloaded files are not deleted. They stay in %LOCALAPPDATA%\QuillVille\Dictation\models (in a portable copy, data\dictation\models inside the portable folder), in folders named whisper-small.en-int8, whisper-distil-small.en-int8, whisper-medium.en-int8, whisper-distil-medium.en-int8 and whisper-distil-large-v3-int8. Delete the folder to free the space.

English only, or many languages? Whisper base.en knows only English and is a little more accurate in English than Whisper base. The many-language ones also dictate Spanish.

Tested on real speech

The advice in QUILL, QUILL Lite and QUILL Transcribe comes from tests with recordings of real people, run on 22 October 2026. Before that it rested on ten recordings of a computer voice, which flattered every engine.

Each model heard public-domain audiobooks catalogued after every model's training cutoff, so no model can have learned them, plus Google's FLEURS read sentences and the LibriSpeech audiobook test sets. It heard them one phrase at a time on one processor thread, the way dictation does, leaving the rest of the computer for the screen reader. Decisions lean on the held-out audiobooks, because some models may have trained on the public sets.

In the tables, "words wrong" is the percentage of words a model got wrong; lower is better. "Speed per thread" is seconds of computing per second of speech on one thread, and dictation's budget is 0.25. "Peak memory" is the most memory the model used on any set. "Not run" means the model was not tested on that set. It was one computer on one day, and other work was running during some sets, so the speeds vary a little.

Moonshine base used to write nothing at all for a phrase longer than about eight and a half seconds. It now hears long phrases in pieces, and scored 5.9% on the held-out audiobooks. The pieces overlap by half a second, so a word that falls where two pieces meet is not lost. Whisper medium and Whisper large-v3-turbo were too slow to test again, and a real 4 GB computer with a screen reader running still needs testing by a person.

The full record of the real-speech tests, 22 October 2026 lists every recording set and its licence, the computer, the limits, and the data behind each row.

How accurate they are

These are the figures each model's makers publish. A word error rate is the share of words a model gets wrong on a recorded test set; lower is better. Test sets differ, so compare numbers only within one line.

• Nemotron 3.5 ASR Streaming: 7.91 per cent on FLEURS English and 4.11 per cent on FLEURS Spanish, at the chunk size this download uses (NVIDIA's model card).
• Parakeet Unified: 1.63 per cent on LibriSpeech test-clean, 3.11 on test-other (model card).
• Parakeet TDT v3: 4.85 per cent on FLEURS English, 3.45 on FLEURS Spanish (model card).
• Whisper on LibriSpeech test-clean: tiny 7.6, base 5.0, small 3.4, medium 2.9 per cent; tiny.en 5.6, base.en 4.2. On Spanish (Common Voice 9): tiny 30.3, base 19.6, small 10.3, medium 6.9 (OpenAI's Whisper paper, Appendix D).
• Whisper large-v3, average over short recordings the model was not trained on: 8.4 per cent (Distil-Whisper model card). OpenAI publishes no figure for large-v3-turbo; it says turbo is much faster than large-v3 with a small drop in quality.
• Moonshine: base 10.07 per cent against tiny's 12.66, averaged over the Open ASR Leaderboard's English sets (Moonshine AI).

In plain words: the bigger models cope better with accents, fast speech, noisy rooms, names and long sentences; every one puts in punctuation and capitals; and for Spanish any many-language model beats the built-in tiny by a wide margin.

What we measured ourselves is in Tested on real speech: recordings of real people, one processor thread, and the most memory each model used. Dictation's budget is 0.25 seconds of computing per second of speech on one thread, so your screen reader always has room. The built-in model is still the better choice on a 4 GB or older computer, on battery, or when your screen reader must never slow down.

Nemotron or Parakeet: which should I try?

Both are NVIDIA models of about 0.6 billion parameters, both are free, and both run here on the processor alone. The difference is when they listen.

• Nemotron 3.5 ASR Streaming is built for live dictation: it recognises while you speak, in short slices, for low delay, and writes punctuation and capitals itself, in English, Spanish and many other languages. Its words appear in the status bar and on braille while you talk (Seeing the words while you speak), and each phrase is written when you pause. It decides whether a sentence was a question once it hears the next one, so a full stop it wrote can become a question mark a moment later. About 682 MB.
• Parakeet recognises each finished phrase after you pause. Nothing appears until you stop, but having the whole phrase can give slightly more accurate final text in some conditions, question marks included. Parakeet TDT 0.6B v3 knows English, Spanish and 23 other European languages (670 MB). Parakeet Unified 0.6B is newer and English only; NVIDIA trained it to work both while you speak and on a finished phrase, and QUILL uses it on the finished phrase (663 MB).

On real recordings, on one thread, Parakeet Unified needed 0.21 seconds of computing per second of speech, Parakeet TDT 0.11 and Nemotron 0.37; at most they used 1.2 GB, 1.1 GB and 1.5 GB of memory. Parakeet Unified got 3.2% of words wrong on the held-out audiobooks, Parakeet TDT 3.8% and Nemotron 5.3%.

Rule of thumb: for the fewest mistakes, try Parakeet: Parakeet Unified for English, Parakeet TDT 0.6B v3 for Spanish. If seeing your words while you still speak matters more, try Nemotron.

How to tell which is more accurate for you

1. Download one model, choose it, and dictate as you normally do for a day.
2. Note each phrase that comes out wrong. Recent Phrases (Shift+F11) shows what was written.
3. Switch to another model in the Speech engine list and do the same.
4. Keep the one that made fewer mistakes, and remove the other.

For an exact answer, QUILL's benchmark measures any model on your own recordings: ten 16 kHz mono WAV files and a transcripts.txt of what you said, then python scripts/dictbench.py folder --model nemotron in a copy of QUILL's source. It reports the share of words wrong, the marks written, the speed, the memory and the load time.

Everything you can say

There are 78 punctuation marks and layout phrases, and 54 spoken commands. This list comes straight from dictation itself, so it is always complete. While you are dictating, say "what can I say" to hear the same list in a window.

What you can say, and what it writes

Punctuation, brackets, symbols and layout. Each table lists the words to say and what appears in your document.

Punctuation

Say any of these anywhere in a phrase. Spaces go where they belong: none before a comma, one after a full stop.

Brackets

Both halves are separate phrases, so you can say one without the other.

Dashes and joining

What "dash" writes is up to you, in Dictation Settings. This row shows what it writes until you change it.

Symbols

The symbols people dictate most often in ordinary prose and in code.

Lines and layout

These change where the next words go rather than adding a character.

Markdown and code

For Markdown and code. A backtick opens or closes depending on how many are already on the line, and a code fence goes on a line of its own.

Starting a line (say these first)

These work only at the start of a phrase, so a sentence that happens to contain the words stays words. If the cursor is not at the start of a line, a new line is started first.

Commands dictation obeys

A command works only when it is the whole phrase, said on its own after a pause. "Delete that line of text" in the middle of a longer sentence is just words, and is written as words. Say "literal" before any phrase on this page to write it out instead of acting on it.

Correcting

Each of these works only as a whole phrase, said on its own after a pause, and only while the phrase it changes is still exactly as dictation wrote it.

Moving the cursor

Said on their own, after a pause.

Dictation itself

Said on their own, after a pause.

Selecting

Say the words you want, and dictation finds the nearest place they appear. If they are not there, your phrase is written as ordinary text and you are told.

Capitals and spacing

Each one lasts until you turn it off or stop dictation, and the status bar shows it while it is on.

Clips, snippets and copying

A snippet or a clip goes in as one phrase, so one Ctrl+Z or "scratch that" takes it back out.

Profiles and formatting

Spelling a word out

• Say "start spelling", then letters. Everything you say is written as letters until you say "stop spelling".
• Say "capital" before a letter for a capital, and "space" for a space.
• Letter names work ("bee", "see"), but the phonetic alphabet works much better: alpha, bravo, charlie, delta, echo, foxtrot, golf, hotel, india, juliet, kilo, lima, mike, november, oscar, papa, quebec, romeo, sierra, tango, uniform, victor, whiskey, x-ray, yankee, zulu.
• Numbers are written as digits: zero, one, two, three, four, five, six, seven, eight, nine.
• Punctuation goes in with no spaces while you spell, and each phrase joins the letters before it, so you can spell an email address a few characters at a time: "jay dot sierra mike india tango hotel", then "at sign echo x-ray dot charlie oscar" writes j.smith@ex.co. A whole word that is not a letter is left out. "Dot" and "point" are a full stop only while spelling.
• Say "all caps" for capitals until you say "no caps" or the phrase ends.
• To spell just one word, say "spell" and the letters: "spell bravo alpha delta" writes bad. A sentence that only starts with the word "spell" is still written as words.
• Say "spell that" to select the phrase you just said. The next phrase is spelled and takes its place.

Starting and stopping without the keyboard

• With the wake phrase switched on in Dictation Settings, say "Quill dictate" to start dictation without touching a key. Anything you say after it in the same breath is written.
• Say "stop dictation" on its own, after a pause, to stop. With the wake phrase on, stopping goes back to waiting for the wake phrase, so you need never touch the keyboard at all.
• You choose both phrases yourself, and each needs at least two words, so ordinary talk cannot start or stop dictation by accident.

Correct that: choosing another guess

Sometimes the engine had a second idea about what you said. Say "correct that" straight after the phrase, and you hear up to three other guesses, numbered: "1: right to Ann. 2: write two Ann. Say choose and the number." Say "choose one", "choose two" or "choose three", and that guess takes the place of what was written. Only Windows speech recognition keeps other guesses; the other engines give one answer each, and "correct that" tells you so. Then say "scratch that" and say it again, or add a correction in My Words and Phrases.

Moving and selecting by voice

You can move the cursor and select text just by saying the words you are after. Say these at the start of a phrase, after a pause.

• "select" and the words. "Select the cat" selects the nearest place those words appear, looking back from the cursor first and then forward. Whatever you say next takes their place.
• "select" one set of words "through" another. "Select dear Sam through thank you" selects everything from the first words to the end of the second.
• "go to" or "go before" and the words puts the cursor just before them. "go after" puts it just after.
• "correct" and the words selects them, and you hear "Correcting the cat. Say the new words". Then say what should be there.
• "select again" or "select next" moves to the next place the same words appear, and "select previous" to the one before.
• Said on their own: "select sentence", "select line", "select paragraph", "go to start of paragraph" and "go to end of paragraph".

Capitals, punctuation and accents do not matter when dictation looks for your words, and numbers match either way, so "two" finds 2. Your corrections in My Words and Phrases are applied first. You hear once where you landed: "Selected: the cat", "Before the cat" or "After the cat". If that is three lines or more from where you were, the line number is added: "Selected: the cat, line 40".

If the words are not in the document, nothing moves. Your phrase is written as ordinary text, and you hear "Not found, written as text." Say "scratch that" to take it back out. All of this works with every speech engine, and the exact commands such as "select that" and "go to top" still do what they always did.

Letters, numbers, symbols and Markdown

Every mark you can say is in the tables above. A few ways of writing are worth a word of their own.

Capitals and spacing

• "caps on" starts every word with a capital until you say "caps off". You hear "Caps on." "Capitals on" and "capitals off" work too.
• "all caps on" writes every letter as a capital until "all caps off".
• "no space on" joins your words with no spaces and no automatic capitals, which is handy for web addresses and file names, until "no space off".
• Each lasts until you turn it off or stop dictation, and the status bar shows which one is on.

Spelling

Spelling mode now writes punctuation with no spaces, so "jay dot smith at sign example dot com" gives you an email address. To spell a single word without switching spelling mode on, say "spell" and the letters: "spell bravo alpha delta". If dictation got a word wrong, say "spell that": the phrase is selected, and the next thing you say is spelled out in its place. There is more in Spelling a word out.

Markdown and code

• "backtick" (or "back quote") opens or closes inline code, depending on how many are already on the line. "Code fence" (or "triple backtick") writes three backticks on a line of their own.
• "tilde", "vertical bar" (or "pipe symbol"), "caret", "greater than sign" and "less than sign" write those characters.
• At the start of a phrase, "bullet" or "list item" starts a bulleted line, "numbered item" a numbered one, "block quote" a quotation, and "heading one" to "heading six" a heading at that level. If the cursor is not at the start of a line, a new line is started first. In the middle of a sentence these stay words, so "the heading two lines down" is written just as you said it.

Snippets and clips by voice

Say these on their own, after a pause:

• "copy all" (or "copy everything" or "copy document") copies the whole document, just like Copy All.
• "copy that" copies the last phrase you dictated.
• "show clips" (or "show copy tray" or "open copy tray") opens the Copy Tray: the Paste from Tray list in QUILL Lite, and the Copy Tray window in QUILL.
• "show snippets" opens your snippets: Snippets in QUILL Lite, and Insert Snippet's list in QUILL.

And start a phrase with these:

• "paste clip" and a number from 1 to 12. "Paste clip three" writes what is in slot 3 of the Copy Tray; "paste slot three" works too. If the slot is empty you are told so, and nothing is written.
• "insert snippet" and its name. You do not need the exact name: capitals, punctuation and spaces do not matter, and the start of the name or one word in it is enough when only one snippet matches. If several match, you hear how many and the list opens so you can choose. If none match, you hear "No snippet called" and the name, and nothing is written.
• "insert abbreviation" or "expand abbreviation" and the abbreviation writes what it stands for.

A snippet or a clip goes in as one phrase, so one Ctrl+Z or "scratch that" takes it out again. The read-back just names it, "Snippet sign off" or "Pasted slot 3", rather than reading the whole thing. In QUILL, a snippet with blanks to fill writes each blank as its name in square brackets, like [name], so you can say "select name" and then say what goes there. None of these work while "Just write what I say" is on, because every command is written as words then.

Dictating in Spanish

New in QUILL Lite 2.0 and QUILL 1.0: you can dictate in Spanish. It is new, so think of it as something to try, and tell us how it goes.

To switch it on, open Dictation Settings (Alt+Shift+F6), move to Dictation language, choose Spanish, and press Enter. Choose English the same way to go back. Quicker still, press Ctrl+Shift+F11 or say "switch to Spanish", as described in Switching between English and Spanish.

What works now

• Your words come out in Spanish, accents and all, using Whisper's multilingual model. It comes with the app, so there is nothing to download. Moonshine only knows English, so in Spanish you get Whisper either way.
• A better Spanish model, said once. The built-in Whisper gets about one word in four wrong in Spanish. If your computer can run NVIDIA Parakeet TDT 0.6B v3 (in practice, 8 GB of memory and four cores), the first time you dictate in Spanish you hear one sentence saying so, for example: "For Spanish, NVIDIA Parakeet TDT 0.6B v3 gets about 6 percent of words wrong, against 29 for the built-in model. It is a 670 MB download in the model list." Nothing is downloaded or changed; the choice is yours. You hear it once, in whichever QuillVille program you dictate in first. A computer with 4 GB hears nothing, because the built-in model is the right one there.
• Windows speech recognition needs Spanish installed in Windows: add Spanish (Spain or Mexico) under Windows Settings, Time and language, Speech. If it is missing, dictation tells you when you start.
• The wake and stop phrases become "Quill dicta" and "deja de dictar". If you typed your own, yours stay.
• Commands stay in English for now. "Scratch that", "select that", "new paragraph" and the rest work as they do in English. Spanish commands come once a native speaker has checked them. The one exception is switching back: "cambiar a inglés" and "dictado en inglés" work already.

Punctuation in Spanish

With automatic punctuation on (the usual setting), Whisper adds the commas, full stops and question marks for you, so just talk. The Spanish punctuation words only work when automatic punctuation is off, or with Windows speech recognition, because "coma" and "punto" are everyday words too. When they are on, you can say:

• "coma" for a comma, "punto" (or "punto y seguido") for a full stop, and "punto y aparte" for a full stop and a new paragraph
• "punto y coma", "dos puntos" and "puntos suspensivos"
• "abrir interrogación" and "cerrar interrogación", and "abrir exclamación" and "cerrar exclamación"
• "abrir paréntesis", "cerrar paréntesis", "abrir comillas", "cerrar comillas"
• "guion", "guion largo" and "arroba"
• "nueva línea", "nuevo párrafo" and "tabulador"

Say "literal" first to write the word itself: "literal coma" writes coma. Say "what can I say" while dictating to hear the full list. If a word keeps coming out wrong, use Help > Get Help from Support and tell us what you said and what was written.

Switching between English and Spanish

Press Ctrl+Shift+F11 (Switch Dictation Language) to change between English and Spanish, even in the middle of dictating. Or just say it: "switch to Spanish" or "Spanish dictation" while you are dictating in English, and "cambiar a inglés" or "dictado en inglés" while you are dictating in Spanish. The accents are optional.

You hear "Español." or "English." and carry on talking. The language you switch to is saved, so next time dictation starts in the one you used last. If you use Moonshine, the first switch to Spanish takes a few seconds while Whisper gets ready; after that, switching back and forth is quick until you close the program.

Dictation never guesses the language for you. It writes in the one you chose until you switch, so a French word in an English sentence does not send it off in the wrong direction.

Teaching dictation your words

Press Alt+Shift+F10, or the My Words and Phrases button in Dictation Settings. A window lists everything dictation has been taught, one line each, and the next phrase you say uses whatever you change. Each change is saved the moment you make it and read back to you. There are three kinds.

• Words Names, jargon and acronyms, spelled the way you want them. When the engine writes something that sounds or looks close, it is corrected to your spelling: add Tucson and "tuxon" becomes Tucson. The match is by spelling and by sound, so split names, run-together words and "R and D" heard for R&D are all repaired, while ordinary words are left strictly alone.
• Phrases Something short you say that writes something longer. Add Phrase asks what you will say (my email address) and what it writes. Press Enter in the second box for a new line, so a signature can be two lines.
• Corrections What the engine keeps hearing wrong, and what to write instead: quill light becomes QUILL Lite. Applied to every phrase, whatever the capitals, on every engine.

Edit (or Enter on a line) changes the line you are on and Remove takes it out. Open the File opens the same list as a plain text file, dictation.md, if you would rather edit a file. Your own phrases appear in the "what can I say" list with everything else.

In QUILL, Locked Dictation and Dictate (Offline) use the same list, so your words are spelled your way there too.

Starting and stopping by voice

The wake phrase

Switch on Listen for the wake phrase while dictation is off in Dictation Settings and you can start dictation by saying it instead of pressing a key. It is "Quill dictate" unless you choose another.

• Say the wake phrase, pause, and start talking. Or say it and keep going: "Quill dictate, dear Sam, thank you for your letter" wakes dictation and writes Dear Sam, thank you for your letter.
• It only counts at the start of what you say, so "I told Quill to dictate this" in conversation starts nothing.
• A near miss still works: "Quil dictate" wakes it.
• Your own phrase needs at least two words. A single word is refused, because it would start dictation by accident.

What it means for the microphone, exactly. While the wake phrase is on, the microphone is open whenever the program is the window in front. Nothing it hears is written, kept or sent anywhere unless it begins with the wake phrase; what does not is dropped the moment it has been checked. When another program comes to the front the microphone closes, and dictation stops if it was running. It does not listen while the program is in the background, so it will never pull you away from what you are doing in another program. The wake phrase is off until you turn it on.

It works with Moonshine, Whisper and Windows speech recognition, but not with Windows voice typing, which Windows runs itself.

The stop phrase

Say "stop dictation" on its own, after a pause, and dictation stops. You can choose your own words as well, in Dictation Settings; "stop dictation" keeps working either way.

• It counts only when it is everything you said in that breath. "I will stop dictation for today" in the middle of a paragraph is written, not obeyed.
• Like the wake phrase it forgives a near miss and needs at least two words.
• With the wake phrase on, stopping goes back to listening for the wake phrase, so you can start and stop without touching the keyboard at all.

Longer pauses, and sentences that carry on

If you like time to think between phrases, open Dictation Settings and set Pause before a phrase is written to Longer (2 seconds) or Longest (3 seconds). Each phrase then waits that long after you stop talking before it is written. Your words appear a little later, but a thinking pause no longer cuts you off.

Dictation also notices when you have not finished a sentence. If a phrase ends on a word a sentence hardly ever ends on, such as "a", "the", "of", "to", "and", "my" or "very", the full stop the engine put there is taken back out and your next phrase carries on the same sentence. Say "this is a very", pause, then say "good test", and you get This is a very good test. A full stop you said yourself is never taken back.

Dictation Settings, every option

Alt+Shift+F6. The window has two columns: what is heard and written on the left, starting and stopping on the right.

Three buttons sit below them: Dictation Commands opens the full spoken list, My Words and Phrases saves these settings and opens the window described above, and More Dictation Settings (Alt+A) opens the window below. The Speech Models and Speech Providers buttons are described in the table above.

More Dictation Settings, every option

The choices nobody needs on the first day. What you change here is saved when you press OK in Dictation Settings.

Knowing what dictation is doing

• The status bar has a Dictation part: Off, Listening, Hearing you, Writing, Spelling, or Waiting for wake phrase. Press F6 to reach the status bar and move to it; Enter there starts or stops dictation.
• The menu row is checked while dictation is writing.
• The tones and the words say when it starts and stops.

Dictation Status: asking what it is doing

Press Alt+F9 at any time to hear what dictation is doing, without changing a thing. With dictation off you hear something like "Dictation is off. Moonshine, in English." With it on you hear the engine and the language, and whether caps or no space is on: "Dictation on, Whisper, in Spanish. Caps on." During a live transcript you hear how long it has been running and how many words it has.

In QUILL Lite it is Dictation Status in Tools, Dictation. In QUILL, Alt+F9 answers for Locked Dictation while that is recording, and for live dictation the rest of the time.

Seeing the words while you speak

With Nemotron (an optional download), OpenAI dictation, or a speech provider that works live (all of them except Google, Groq and Mistral), the program knows your words while you are still saying them. The status bar shows them after the word "Hearing", and a braille display shows them as they arrive. They are not in your document yet, because the engine may still change its mind; when you pause, the final words go in and the preview goes away. Nothing is spoken, so nothing talks over you. To hear new words quietly as well, choose Show it, and say new words quietly under While you speak, the words heard so far in More Dictation Settings.

Words go where you started

Into the document you started dictation in, at the cursor, replacing any selection. Your words go where you were when you started speaking: if you move the cursor, or switch to another window, while a phrase is still being recognised, it is still written there, your cursor is put back where you moved it, and you hear "Written where you started". If that spot changed meanwhile, the words go at the cursor and you are told. After that, dictation stops by itself when that document closes, when you switch it between plain and rich text, or when you speak while another program or a dialog is in front, and the program tells you why it stopped.

There is one microphone, so pressing Ctrl+F11 in a second document while dictation runs in the first moves dictation there and says so: "Dictation moved to Letter.txt". Dictation will not start in a read-only document, such as the read-only copy Compare opens, and says so before the microphone is opened.

If the microphone drops

If the microphone is unplugged, a Bluetooth headset drops, or the input goes completely silent for a few seconds, dictation says "The microphone stopped. Dictation is paused and will resume when it comes back." It stays on, writes nothing, and watches for the device; when it returns you hear "Microphone back. Listening." and carry on. The status cell says "paused, microphone lost" meanwhile.

If the speech engine itself stops answering it is restarted once, silently, and you notice nothing. Only a second failure is reported, by name, with the engine to try instead: "Moonshine stopped working. Try Whisper in Dictation Settings."

Dictating into Find, Replace and the AI box

Ctrl+F11 in the Find box, either Replace box, the question box of the AI pad, or the message box of the AI Conversation window dictates into that box: the same engine, the same words you have taught it, and Escape still cancels a phrase. "New line" and "new paragraph" become a space in a one-line box, no closing full stop is added, and the commands that move around a document do nothing there. Press Ctrl+F11 again to stop.

Dictating into other programs with Quill Inkwell

Dictate Anywhere lets you dictate into your email, a web page, a chat window or any other program, with the same engines, the same words you have taught it and the same voice commands. It lives in Quill Inkwell, the free family app that already sits quietly in the background and types into other programs for you.

1. Open Quill Inkwell and choose File, Dictate Anywhere Key. Press the key you want to use anywhere in Windows. There is none until you choose one, so it can never clash with a key you already use.
2. Go to the program you want to write in, and put the cursor where the words should go.
3. Press your Dictate Anywhere key, speak, and pause. Each phrase is typed where the cursor is.
4. Press the key again, or say "stop dictation", to stop.

The easy way to start: in QUILL or QUILL Lite, open More Dictation Settings and press Dictate in Other Programs. Your dictation settings are handed to Quill Inkwell and it starts; if you have not chosen a key yet, the key chooser opens for you. After that, Inkwell keeps its own copy of the settings, which you can change in its Dictation menu under Dictation Settings. It uses the same My Words and Phrases list as the editors.

Punctuation, spelling, caps and no space, switching language and "scratch that" all work in other programs. "Scratch that" erases what was just typed, but only while you are still in the same window, so it can never erase somebody else's typing. Anything that needs to read the other program's text, such as selecting by voice, "go to", "correct", clips and snippets, cannot work there, and you hear "That works in QUILL's own documents. In another program, say the words again, or scratch that."

To keep things safe, nothing is typed into a password field ("That is a password field, so nothing is typed there."), into a program running as administrator when Inkwell is not, or into Inkwell's own window. In QUILL's own windows you are reminded to use Ctrl+F11, which already dictates there. The read-back starts as a tone only, because your screen reader already echoes what is typed. In Safe Mode, Dictate Anywhere is off.

QUILL Transcribe, releasing 1 November 2026

QUILL Transcribe is a free family app made only for dictating into other programs, and it is released on 1 November 2026. It lives in the system tray, uses the same speech models, words and commands as QUILL and QUILL Lite, writes only into the field you started in, and pauses rather than typing somewhere else if you move. Its start and stop key is Alt+Shift+Backslash unless you change it. While it is running it does the dictating for Dictate Anywhere too, so only one of the two listens at a time.

For example, to answer an email by voice: open QUILL Transcribe once and choose a speech model, then go to the body of your reply in your email program, press Alt+Shift+Backslash, say "thanks for the update period I will call you tomorrow period", and press Alt+Shift+Backslash again.

Every other released family app opens it from its QuillVille menu, and Quill Cast and Quill Radio use it for Edit, Dictate Here (Ctrl+F11), which dictates into their search boxes. There is no download yet; the QUILL Transcribe user guide describes what it does.

Talking to the AI

You can have a spoken conversation with AI help. Open the AI Conversation window; the cursor is already in Your message.

1. Press Ctrl+F11 and ask your question.
2. Pause. Your message is sent, and you hear "Working."
3. The reply is read aloud. While it is read, the microphone does not listen, so the reply is never taken as your next message.
4. When it has been read, carry on talking. To speak sooner, press Escape: you hear "Listening."
5. Press Ctrl+F11 again when you are done.

In that box dictation uses its Talking to AI settings by itself: a longer pause, so a breath does not send half a question, filler words removed, and automatic punctuation on. They are in More Dictation Settings, with Send a dictated message to the AI, which you can set to When I press Enter to check each message first.

OpenAI dictation, with your own key

Optional, and off until you choose it. Your computer's own engines stay the default.

With your own OpenAI key you can have OpenAI recognise your speech. It is very accurate, punctuates well, gets names right that it has never met, and can show your words as you speak. What you say is sent to OpenAI: each phrase goes over an encrypted connection with your own key, OpenAI bills your account, nothing goes through QUILL's servers, and QUILL keeps no copy of your voice. Silence is never sent. QUILL's free AI is never used for dictation, and it does not work in Safe Mode.

1. Save your OpenAI key in Use My Own AI Key (Alt+F2), or from Dictation Settings: More Dictation Settings..., then Add or Change OpenAI Key... (Alt+K).
2. Open Dictation Settings (Alt+Shift+F6) and choose OpenAI (your own key; sends your speech to OpenAI) as the speech engine.
3. A question says exactly what is sent. Choose Yes to agree; No, the answer if you just press Enter, keeps your engine and sends nothing.
4. More Dictation Settings opens on OpenAI speech model, listing the models OpenAI offers your key, newest first, with the newest chosen. Models OpenAI is retiring are left out. gpt-live-transcribe writes as you speak; gpt-transcribe sends each phrase when you pause.
5. Press Enter twice to save, then dictate with Ctrl+F11 as before.

If OpenAI stops offering your model, you are told once and asked to choose another; dictation never changes it for you. The words in My Words and Phrases are sent as words to expect, so names come out right. To stop sending your speech, choose another engine, or turn off Let OpenAI dictation send my speech to OpenAI in More Dictation Settings.

Speech providers, with your own account

Optional, and off until you choose one. If you already pay for a cloud speech service, dictation can use it. The engines on your computer stay the default and the recommendation.

QUILL, QUILL Lite, QUILL Transcribe and Quill Inkwell can each send your speech to one of twelve speech services, using an account you already have with that company. You enter only what that service needs, such as a key and a region. The company bills you, not QUILL, and QUILL shows no prices. Nothing goes through QUILL's servers, and QUILL keeps no copy of your voice. Why choose one: a service you trust and already pay for, very good accuracy on a computer too modest for the big local models, or a service tuned for your language or field. Why not: your speech leaves your computer, it needs the internet, and it costs money.

Setting one up

1. Open Dictation Settings (Alt+Shift+F6; in Quill Inkwell, Ctrl+Alt+D) and press Speech Providers... (Alt+V). Focus starts in the list of providers.
2. Arrow to your provider. About this provider says what it offers and where your speech goes.
3. Tab to its fields and fill them in: a key, and for some a region, a project or a file. Press F1 on any field to hear where to find that value in your account.
4. Press Save (Alt+S). A key goes straight into Windows Credential Manager, never into QUILL's settings or logs, and a saved key is never shown again. The other fields are saved when you press OK in Dictation Settings.
5. Press Test Connection (Alt+T). It asks the provider about your account and never sends speech. Last result says what happened, such as a key that was refused.
6. Close Speech Providers. Your provider is now in the Speech engine list, as for example "Azure AI Speech (your own account; sends your speech to Microsoft)". Choose it.
7. A question says exactly what is sent, where, and that the company bills you. Choose Yes to agree. No is the answer if you just press Enter: your engine stays as it was and nothing is sent.
8. Press OK, then dictate with Ctrl+F11 as usual.

For example, to use Microsoft's service: in the Azure portal, open your Speech resource and its Keys and Endpoint page, and copy Key 1 and the Location. In Speech Providers, choose Azure AI Speech, paste the key into Key, choose the same region in Region, press Save, then Test Connection. When Last result says it worked, close the window, choose Azure AI Speech in the Speech engine list, read the question, and answer Yes.

What is sent, and what happens when it fails

• Sent: each phrase the voice detector hears, over an encrypted connection, with the account details you saved. Nothing is sent while nobody is speaking. The words in My Words and Phrases go too, as words to expect, where the provider accepts them. The company's own policies apply to what it receives.
• The network drops: that phrase is lost, you hear one sentence saying so, and dictation keeps listening.
• The key is refused or the account is empty: dictation stops and says which field to check.
• Never a silent switch: dictation never changes to another engine by itself. To stop sending your speech, choose another engine in Dictation Settings.
• Remove (Alt+M) in Speech Providers forgets everything saved for that provider, key included.
• Safe Mode refuses both Test Connection and dictating with a provider.

Limits for now. The providers are for live dictation; Transcribe a Recording offers only the models on your computer and OpenAI. QUILL Transcribe's Choose Model window does not list providers yet, so choose them in its Dictation Settings (Alt+Shift+F6). Google has no live words while you speak. And the providers have been tested against recorded answers but not yet against real accounts, so please tell us how yours goes.

Live transcripts

A live transcript writes down everything you say, for as long as you like, in a document of its own. It is made for a meeting, a lecture, or a long stretch of thinking out loud. Press Ctrl+Alt+Shift+PageDown (Start or Stop Live Transcript), or choose it from the dictation menu.

• A new untitled document opens, and you hear "Live transcript on, in a new document." It never writes into the document you were working on.
• Everything is written down as it is: no voice commands except the stop phrase, no tone or read-back after each phrase, filler words left out and punctuation added for you. It never stops by itself after a silence.
• A pause of four seconds or more starts a new paragraph. Check Time stamps in live transcripts in More Dictation Settings and each paragraph starts with the time, like [10:42].
• New words always go at the end of the transcript, even while you work in another document or another program, and your own place in the transcript is kept, so you can read back over it while it carries on.
• The status bar shows how it is going, as in "Live transcript: 12 minutes, 1,840 words", and Alt+F9 says the same.

To stop, press Ctrl+Alt+Shift+PageDown again, or Ctrl+F11. You hear "Live transcript stopped" and the number of words, such as "Live transcript stopped, 1,840 words." The document stays open, unsaved, for you to save where you like; until you do, it is protected like any other unsaved document. In QUILL you can then run a Transcript Action on it.

The first time you start one, you hear "Please record other people only when they have agreed." Live transcripts listen to your microphone only, not to sound playing on the computer, and they need one of the program's own engines rather than Windows voice typing.

Transcribing an audio file

A recording into text, in the background, with the same speech engines you dictate with. The same command, window and key in QUILL and QUILL Lite.

Sometimes the words you want are already in a recording: a voice note from your phone, an interview, a lecture, a meeting somebody recorded for you. Press Shift+F5 (in QUILL Lite, Tools, Dictation, Transcribe a Recording; in QUILL, Tools, Speech, Live Dictation, Transcribe a Recording), choose the file, and keep working while it listens.

1. Type the path of the recording, or press Browse... and choose it. You can drop a file on the window too, and choose several in Browse to have them done one after another.
2. Check the language: English or Spanish.
3. Check the speech model. The most accurate one on your computer is already chosen, and the box under it says how long the recording is and about how long it should take.
4. Choose where the text goes: a new document, which is already chosen, or this document, at the cursor, when it finishes.
5. Press Enter. You hear when it starts, the percentage quietly at each quarter, and one sentence when it is done, such as "Transcribed meeting.mp3: 12 minutes, 1,804 words, in 6 minutes, in a new document."

Paragraphs start where the recording pauses for two seconds or more. Timestamps are off unless you turn them on: check Add timestamps and each paragraph begins with its time, like [00:01:23]. Spoken "comma" or "new paragraph" are written as words, because in a recording people usually mean the words; check Obey spoken punctuation and commands for a recording you dictated on purpose. Your corrections from My Words and Phrases are used, and filler words go if you remove them when you dictate.

Which model for which recording

• Voice notes and dictated letters, one clear voice: Moonshine tiny, built in, is the fastest and does this well.
• Interviews and meetings, several voices: a downloaded model is worth it. Parakeet TDT 0.6B v3 is our first choice; Whisper small, medium or large-v3-turbo are good too, just slower.
• Lectures and talks: Parakeet or a Whisper model, with the speaker's names added to My Words and Phrases first.
• Noisy rooms: the bigger the model, the better it copes. The two tiny models struggle with noise.
• Spanish: the built-in Whisper tiny, Parakeet TDT 0.6B v3, Nemotron and the multilingual Whisper models.

Files it reads: MP3, M4A, AAC, WAV, Ogg, Opus, FLAC and WMA, plus the sound in an MP4 and AIFF, in both products, with parts of Windows that are already on your computer. Nothing extra to download, and no ffmpeg needed. Stopping: press Shift+F5 again and choose Stop Transcribing. Every result, finished or failed, stays in Activity (Shift+F9).

Privacy. With the models on your computer, nothing leaves it. If your own OpenAI key is saved, OpenAI's speech models join the list, but they are never chosen for you, never offered in Safe Mode, and each recording is sent only after you say Yes to a question that says exactly what is sent. It goes about a minute at a time over an encrypted connection, OpenAI bills your account by the minute, and QUILL keeps no copy. Please only send recordings you have the right to share.

Tidy Dictated Text: tidying with AI, when you ask

Speech recognition writes what it heard, and what it heard is not always the word you meant: their for there, a name it has never met, two words run together, a comma where a full stop belonged, and every "um" you did not know you said. Ctrl+F3 fixes exactly that and nothing else.

1. Dictate as usual. When the paragraph is done, leave the cursor in it, or select exactly the stretch you want tidied.
2. Press Ctrl+F3. You hear "Working." The selection, or else the paragraph the cursor is in, goes to the model with one instruction: correct misheard words, punctuation and capitalisation, remove fillers and false starts, and change nothing else.
3. The Tidied Dictation window opens with focus on the corrected text, so you hear it first. Replace My Selection puts it where the dictated text was, Copy puts it on the clipboard, and Escape keeps what you had. If you typed in that paragraph while the answer was on its way, Replace is not offered and the window says so.
4. Ctrl+Z takes a replacement back, as one step.

Nothing changes until you press Replace, and the model is told to add nothing. It needs a ChatGPT subscription or your own AI key (OpenAI or Google Gemini). QUILL's free AI service does not offer it. If you have neither, it tells you and opens the account window.

My Dictation Instructions

Tell Tidy Dictated Text how you like your writing. In Dictation Settings press More Dictation Settings (Alt+A), then My Dictation Instructions (Alt+I). A short file opens; write one instruction per line, such as "Write numbers as digits" or "Use British spelling", and save it. They go with the passage you tidy, and only then. The AI is told your dictated text is something to correct, never a request to answer.

Dictation context for each document

Tell dictation what a document is about, and OpenAI dictation and Tidy Dictated Text can use that to get your words right. Press Ctrl+Alt+Shift+PageUp (Dictation Context for This Document). The window has four parts:

• This document is: describe it in your own words, such as "a letter to my landlord about the boiler" or "notes from the garden club".
• Start from a saved context: the contexts you have saved, then a few to start from: Formal letter, Note to a friend, Technical writing, Meeting notes and Story.
• Also save it as a context named: give it a name to use it again in other documents.
• Who uses it: a read-only box that tells you which parts of dictation will use it.

Each document remembers its own context. An untitled document keeps it until you close the program. While dictating, say "dictation context" and the name of a saved one, such as "dictation context meeting notes", to choose it by voice.

OpenAI dictation uses the context to understand you better, and Tidy Dictated Text uses it alongside My Dictation Instructions. The other engines, downloaded models and speech providers included, cannot use it, and the window tells you so. Your context only ever goes where your speech or your text is already going.

What leaves your computer

Dictation sends nothing unless you choose OpenAI dictation with your own key, or a speech provider with your own account, and agree to it. One optional command, Tidy Dictated Text, sends a passage when you press it.

Nothing above happens in the background or as you type. Dictation opens the microphone when you ask it to and closes it when you stop. The wake phrase is the one setting that keeps the microphone open. It is off until you turn it on, it closes the microphone whenever another program comes to the front, and it never listens while the program is behind another window.

On a modest computer

Nothing runs while dictation is off: the speech model loads the first time you start, stays ready while you use it, and is put away a few minutes after you stop, giving its memory back. If an optional model you downloaded cannot keep up with your voice, because the computer is busy or on battery saver, you hear "This computer is busy, so dictation switched to the faster built-in engine for now. Your choice in Dictation Settings is unchanged." Next time, it tries your model again.

Dictating well

A few habits make a big difference, whichever engine you use.

• Speak in whole phrases. Say a sentence or a clause the way you would to a person, then pause. Engines understand a whole phrase far better than one word at a time.
• Pause before and after a command. "Scratch that", "select the cat" and the rest only count as a phrase on their own, so a short pause either side is what tells dictation you mean it.
• Need time to think? Choose a longer pause. Longer or Longest in Dictation Settings, so a thinking pause does not end your sentence. If it does, and the sentence ended on a word like "the" or "and", dictation joins it back up for you.
• Fix words by saying them. "Correct" and the wrong words, then the right ones, is often quicker than reaching for the keyboard. "Spell that" is there for a name the engine has never heard.
• Teach it once. A name it keeps getting wrong belongs in My Words and Phrases (Alt+Shift+F10), so you never correct it again.
• Use a headset. A headset microphone beats a laptop's built-in one, and keeps the read-back out of the microphone.
• Lost track? Press Alt+F9 to hear what dictation is doing, or Shift+F11 for the last twenty phrases.

If something goes wrong

Dictation always says what happened and what to do about it. These are the usual reasons, and the fix for each.

It will not start

• No microphone was found. Connect one and press Ctrl+F11 again.
• The microphone you chose is not connected. Plug it in, or choose another in Dictation Settings. It will not quietly listen on a different one.
• The microphone could not be opened. In Windows Settings, go to Privacy and security, then Microphone, and make sure desktop apps are allowed to use it.
• A speech engine is not included in this copy. Choose another engine in Dictation Settings; reinstalling puts the missing one back.
• No speech recogniser is installed, with Windows speech recognition. Add a speech language in Windows Settings, Time and language, Speech.
• "This document is read-only, so dictation cannot write here." Dictate into a document you can type in.

It keeps mishearing me

• Try Whisper instead of Moonshine, or the other way round. They fail on different voices. Or try a downloaded model: Choosing an engine says which.
• Use a headset microphone rather than a laptop's built-in one. This is the single biggest improvement available to most people.
• If whole sentences go missing, your microphone may be too quiet. With its level set very low, or speaking very softly or far away, dictation can miss speech entirely, especially just after something louder; we saw it happen in our tests of very quiet speech. Move closer, or raise the level in Windows' sound settings.
• If you are using speakers, set the read-back to a sound only. Otherwise the microphone hears the read-back and writes it down again.
• Add the names you use often to My Words and Phrases (Alt+Shift+F10), and add a correction for anything it gets wrong the same way every time.
• Choose a Long pause if it cuts you off mid-thought.
• Tidy the result afterwards with Ctrl+F3, if you have a ChatGPT subscription or your own OpenAI or Google Gemini key.

It stopped on its own

• "Moonshine stopped working. Try Whisper in Dictation Settings." The engine failed twice in one session; the first time it was restarted for you without a word. Choose the other engine, and if it keeps happening, write to support.
• After a few minutes of silence. That is "Stop dictation after silence" doing its job; it says so, with the number of minutes. Set it to Never in Dictation Settings if you would rather it waited.
• When you moved to another program. Dictation follows the document, not the screen. It tells you why it stopped.

A command was written as words instead of obeyed

A command counts only as a whole phrase, on its own, after a pause. Said inside a sentence it is just words. That way you can still dictate the sentence "delete that line of text". Check also whether "Just write what I say" is switched on, which writes every command as words by design.

If you hear "Not found, written as text." after "select" or "go to", the words you asked for are not in the document, so your phrase was written instead. Say "scratch that", then try again with words that are there.

I said "scratch that" one time too many

Press Shift+F11 for the last twenty phrases and insert the one you lost. That is what the list is for.

Which product has which dictation

Live Dictation is identical in both, down to the keys. Everything QUILL adds is a second, different way to dictate.

A dashed line down the left of a row means one of the two does not have it. For everything else the two products differ on, see how QUILL and QUILL Lite compare, capability by capability.

Downloads and further reading

Dictation is in both programs at no cost and with nothing to add. In QUILL Lite it is on in every feature profile except WordPad and Notepad, and can be switched off in Customize Features like any other area. Switching it off also removes its menu, its keys and its wake phrase.

Get either program

QUILL Lite 1.1.2 installer for Windows All QUILL 0.9.0 Beta 2 downloads, including macOS and portable

Windows 10 or later, 64-bit for dictation in either program. Dictation only works on Windows, so QUILL on a Mac has everything else but not dictation. QUILL Lite's installer is about 230 MB, mostly because the two speech engines are inside it.

Read more

• The QUILL Lite user guide: dictation alongside everything else the program does.
• The QUILL user guide: Locked Dictation, transcription, captions and voice commands.
• How QUILL and QUILL Lite compare: everything that is different between the two.
• The QUILL Transcribe user guide: dictating into any program, from 1 November 2026.
• The QUILL Lite AI guide: what Tidy Dictated Text sends, and what the free service does.
• QUILL frequently asked questions: installing, updating and privacy.

Back to the top of this page