How to Choose a Dictation App in 2026
2026-10-05 · 69 min read · Yap team
A buyer's guide to dictation apps for Mac and Windows, built on evidence you can check yourself
Speech is about three times faster than typing. In a controlled study of short-message entry on an iPhone, English speech reached 153 words per minute while the built-in keyboard managed 52 words per minute, and Mandarin speech beat Pinyin typing by a similar margin - Ruan et al. on arXiv. Desktop keyboards do not close the gap for most people either: an online study of 168,000 volunteers found that the average participant typed 52 words per minute, with most people between 30 and 60 - Aalto University. A dictation app is the tool that turns that speaking speed into finished text wherever your cursor happens to be.
The hard part is choosing one. Dictation apps look alike in a screenshot: a keyboard shortcut, a microphone glyph, and words appearing where you were about to type. The differences that decide whether you keep using one are invisible at download time. Is your audio recognized on your own computer or uploaded to a vendor's server? What happens in your second language? How long do you wait after you stop talking? Does the text land in every app you use, including the ones with password fields? And what does it cost once the trial ends? Most comparison pages on these questions are written by the vendors themselves, and this one is no exception: we make Yap, an on-device dictation app for Mac and Windows.
So we built this guide to be checked rather than trusted. Every price, policy and capability links to the vendor's own page as we read it on October 5, 2026. Every accuracy and speed number comes from a public benchmark or from our own tests, with the method laid out in section 14. And Yap is scored by the same rules as every other option, which puts it fourth of nine in our scorecard. Where another app is the better choice for your situation, we say so plainly, and we name it.
What this guide covers: what a dictation app actually does between your voice and your cursor, why the place your speech is processed matters more than any feature list, how to read word error rates and test accuracy on your own voice in ten minutes, what the wait after you stop talking is made of, how languages, formatting and AI rewriting change the result, how text gets into other apps safely, and what each option really costs over one and three years. It ends with a decision framework that maps common situations to a short list you can test this week.
Contents
- The 2026 dictation app scorecard
- What a dictation app actually does
- Why dictate at all: the speed math
- Where your voice goes: on-device or cloud
- Accuracy: what word error rate tells you, and what it hides
- Speed: the wait after you stop talking
- Languages, accents and switching between them
- Formatting, voice commands and AI rewriting
- Typing into any app, safely
- What dictation apps really cost
- Mac, Windows or phone: platform coverage
- When the built-in tools are enough
- A decision framework
- How we researched and tested this guide
1. The 2026 dictation app scorecard
Before the detail, here is the field on one page. The table scores nine ways to dictate on a Mac or a Windows PC: the two tools built into the operating systems and seven apps you install. We chose them because each one types into any application rather than into a single document, and because each publishes enough about its pricing and data handling to be scored on evidence rather than impressions. One well-known name is left out of the table on purpose and covered in prose later: Dragon Professional, whose official page lists no price to score.
Every cell gives a score from 0 to 10 and the fact that produced it, so you can disagree with our judgment without redoing our research. The final score is the weighted average, rounded to one decimal, and the table is sorted by it; ties are listed alphabetically. Read the justification before the number: two apps can earn the same score for opposite reasons, and the reason is what tells you whether an app fits your situation.
| # | App | What it does | Privacy (30%) | Engine and languages (25%) | Reach (20%) | Cost (25%) | Final |
|---|---|---|---|---|---|---|---|
| 1 | Superwhisper | Dictation on offline Whisper models or cloud AI | 8 - works offline; cloud models optional | 8 - local Whisper or cloud models, 100+ languages | 9 - Mac, Windows, iOS, Android | 8 - free tier with unlimited Whisper models; Pro $8.49/mo | 8.2 |
| 2 | VoiceInk | Open-source Mac dictation on local models | 9 - transcription stays on the Mac; optional cloud cleanup sends text only | 8 - local Parakeet and Whisper models | 6 - Apple Silicon Macs, new iOS app | 9 - $25 once for one Mac | 8.2 |
| 3 | Apple Dictation | Built into every Mac | 8 - on device for 48 of 63 language entries | 7 - 63 language entries, no published accuracy | 6 - every Mac, any text field; Mac only | 10 - free | 7.9 |
| 4 | Yap | On-device dictation for Mac, Windows, iPhone and Android | 10 - recognition and history stay on the device | 7 - Parakeet v3 family, 25 languages | 8 - Apple Silicon Macs, Windows 10 and 11 (x64 and Arm), iPhone (beta) and Android | 6 - $69/yr or $6.90/mo after 3 free days and a 10-day trial, no free tier | 7.7 |
| 5 | MacWhisper | Mac transcription suite with dictation in any app | 8 - local models on the Mac; cloud models optional | 7 - local and cloud models; languages not listed on its page | 5 - Mac only | 8 - €64 once, lifetime updates; free version | 7.2 |
| 6 | Willow Voice | Cloud dictation with an offline mode | 6 - offline mode runs on device; cloud otherwise | 7 - own models, 100+ languages, no public benchmark | 8 - Mac, Windows, iPhone | 6 - free on a weaker model; Pro $12/mo billed yearly | 6.7 |
| 7 | Windows voice typing | Built into Windows (Win+H) | 4 - online Azure speech; recordings not kept without permission | 7 - Azure Speech, 46 language entries | 6 - every Windows PC, any text field | 10 - free | 6.7 |
| 8 | Wispr Flow | Cloud dictation with AI formatting | 3 - always cloud; data can train models if sharing is on | 7 - cloud models, 100+ languages, no public benchmark | 10 - Mac, Windows, iOS, Android | 4 - 2,000 free words a week; Pro $12/mo billed yearly | 5.7 |
| 9 | Aqua Voice | Cloud dictation on its own Avalon model | 3 - cloud only; transcripts may be kept unless Privacy Mode is on | 7 - Avalon model, 49 languages | 8 - Mac, Windows, iOS; Android waitlist | 5 - 1,000 free words; Pro $8/mo billed yearly | 5.5 |
How the four criteria work. Privacy (30%) asks where your speech is recognized and what the vendor says it keeps: a 10 means nothing leaves the computer, low scores mean every word is processed on a server and may be stored or used to train models. Engine and languages (25%) rewards public, checkable evidence about the recognition model and the number of languages it handles. It is deliberately not a measurement of each app on your voice: no public test runs all nine on the same recordings, and we would rather say so than invent one. Section 5 shows how to run that test yourself in ten minutes. Reach (20%) counts platforms and whether the app types into any application. Cost (25%) reflects what a year of full use costs, with free tiers counted when they are genuinely usable every day.
Why privacy weighs the most. It is the hardest property to change after you commit. A slow app can get faster in an update and a missing language can be added in the next release, but an app whose design sends every word to a server is a business decision you inherit, together with whatever its privacy policy says next year. Speed is not scored because vendors publish targets rather than comparable measurements; section 6 covers what we could actually verify, including our own timings on two different Macs.
The result is less flattering to the best-known names than their marketing suggests, and less flattering to us than you might expect from a vendor's guide. The two apps at the top combine local recognition with low prices, and the free Apple Dictation beats several paid apps once cost is counted. Yap lands fourth: it has the highest privacy score, but it costs money once its 3 free days and its trial are over, and its phone apps are newer than its Mac and Windows apps (an iPhone beta and an Android app). The cloud apps at the bottom are not bad products, and Wispr Flow reaches more platforms than anything else in the table; they rank low because privacy and cost together carry more than half the weight.
Why this matters: a scorecard can only rank what you care about if its weights match yours, so treat it as a starting shortlist, not a verdict. How to apply it: if you need the same dictation on your phone, your real shortlist is Superwhisper, Wispr Flow, Aqua Voice, Willow Voice and VoiceInk's new iOS app, whatever their privacy scores. If you only use a Mac and want to pay nothing, start with Apple Dictation. Either way, take your top two or three through the accuracy test in section 5 on your own voice before paying for anything.
2. What a dictation app actually does
Every dictation tool, from the one built into macOS to a subscription service, runs the same four steps. It captures audio from your microphone, recognizes speech as words, formats those words into text, and inserts the text into the app you are using. Each step can fail on its own, and most disappointment with dictation comes from one weak step rather than from recognition as a whole. A brilliant recognizer is useless if the first word of every sentence is clipped because recording started late, or if the finished text lands in the wrong window because you clicked somewhere else while you were still talking.
The step that separates apps most sharply is the second one, because it decides who hears you. Recognition is the heavy computation: NVIDIA's Parakeet TDT 0.6B v3, for example, has 600 million parameters and takes in 16 kHz audio, meaning 16,000 samples for every second you speak - NVIDIA model card. That work can run on the chip in your own computer, or your audio can be uploaded to the vendor's servers and the text sent back. Almost everything else in this guide, from privacy to speed to price, follows from where that one step runs.
Capture sounds trivial and is not. Push-to-talk apps record while you hold a key, toggle apps record between two presses, and hands-free modes decide you are finished when they hear silence. Apple's own Dictation, for example, lets you dictate text of any length but stops automatically when it detects no speech for 30 seconds - Apple Support. The practical risks are a clipped first syllable when recording starts a beat after your key press, a microphone the app did not switch to when you plugged in a headset, and recordings cut short by a time limit. Long thinking pauses are normal when you compose by voice, so check how an app treats them before you trust it with a long email.
Recognition turns audio into words with a trained model, and two open model families dominate local dictation in 2026. Whisper, released by OpenAI under the MIT license, comes in six sizes from tiny to large, including a turbo model that the project describes as "an optimized version of large-v3 that offers faster transcription speed with a minimal degradation in accuracy" - OpenAI Whisper on GitHub. Parakeet TDT 0.6B v3 covers 25 European languages, detects the spoken language by itself and adds punctuation and capitalization in the same pass. Cloud apps run their own models, such as Aqua Voice's Avalon, or models they license from others, and they rarely publish which.
Formatting turns a stream of words into readable text: punctuation, capital letters, numbers written as digits, paragraph breaks, and sometimes the removal of filler words like "um". Some of that comes from the recognition model itself, the rest from rules, from spoken commands such as "new paragraph", or from a language model that rewrites what you said, which section 8 covers in detail. Insertion then puts the result where your cursor is. An app can simulate keystrokes, paste through the clipboard, or use the operating system's accessibility interface. On a Mac, an app that controls other apps needs your approval under Accessibility in Privacy and Security settings - Apple Support, and on Windows, simulated input can only reach applications "at an equal or lesser integrity level" than the app sending it - Microsoft Learn.
The diagram's last box, history, is optional and many apps skip it, but it changes how failures feel. If an app saves each result before it tries to insert it, a failed insertion costs you a click instead of a paragraph: you open the history and copy the text. If it does not, anything that goes wrong after recognition, from a closed window to a refused paste, means speaking the whole thing again. Where that history is stored is a privacy question in its own right, and section 9 returns to it.
Four quick checks expose a weak step faster than any feature list:
- Capture: start talking the instant you press the key; does the first word survive?
- Recognition: turn off Wi-Fi; does it still work, and how fast?
- Formatting: can you get exactly what you said, without rewriting?
- Insertion: click into another window while talking; where does the text go?
Each check maps to a common way dictation fails in daily use, and none of them shows up in a vendor's demo video, because demos are recorded on fast connections by people who know exactly when to start talking. The second check is the most revealing one, because it tells you in a few seconds whether the privacy claims are about architecture or about policy. The fourth check matters more than it sounds: a dictation that lands in a chat window instead of the document you meant can send half a thought to the wrong person, and the only real protection is an app that notices the change or keeps a copy you can recover.
Why this matters: when dictation disappoints, people blame "accuracy" and switch apps, when the real problem was capture or insertion and the new app has the same weakness. How to apply it: run the four checks above on any app during its first ten minutes, before you look at its accuracy at all, and drop any app that fails the capture or insertion check, because no amount of recognition quality makes up for words that are lost before or after the model sees them.
3. Why dictate at all: the speed math
The case for dictation is arithmetic, and the arithmetic has two sides: how fast words go in, and how long fixing them takes afterwards. The input side is the well-known one. In the study by Ruan and colleagues, participants transcribed short messages on an iPhone 6 Plus, either with Baidu's Deep Speech 2 speech recognizer or with the built-in iOS keyboard, under laboratory conditions for both methods - arXiv. Speech came out 2.93 times faster in English and 2.87 times faster in Mandarin Chinese, a gap large enough to survive any reasonable doubt about lab conditions.
Desktop typing is not dramatically different for most people. The largest study of everyday typing, run by researchers at Aalto University and the University of Cambridge, measured an average of 52 words per minute; professionally trained typists in the 1970s and 1980s typically reached 60 to 90, and the fastest participants in the new study reached 120 words per minute - Aalto University. Even a very fast typist, in other words, is slower than an ordinary speaker reading a message aloud.
The correction side is quoted less often and matters more. In the same study, speech made fewer errors during entry (a corrected error rate of 5.30% against 11.22% for the keyboard) but left slightly more errors in the final text (1.30% against 0.79%). That pattern is the whole story of dictation in two numbers: speaking is much faster at getting words out, and the remaining cost lives in finding and fixing the words the recognizer got wrong. The speed advantage is real, but it is a budget, and recognition errors spend it.
Here is a simple model that makes the budget concrete. Take a 300-word email. At the average typist's 52 words per minute it takes about 5.8 minutes to type. Spoken at 153 words per minute, the words are in after about 2 minutes, which leaves roughly 3.8 minutes for fixing mistakes before typing would have been faster. If each fix costs you about 5 seconds, you can afford about 45 fixes, so dictation stays ahead until roughly 15 percent of the words are wrong. If each fix takes 10 seconds, because you reach for the mouse, find the word and retype it, the break-even falls to about 7.6 percent, or one word in thirteen.
This is a model, not a measurement: the study's speeds already include corrections made during entry, and your own pace depends on how you compose. But it explains something people notice without being able to name it. With a recognizer that gets 4 or 5 percent of your words wrong, dictation feels like a superpower; at 12 or 15 percent, it feels like a wash, and people quietly go back to the keyboard. The difference between those two experiences is rarely the app's speed. It is accuracy on your voice, in your room, with your vocabulary, which is why section 5 spends so long on how to measure it.
There is also a human side to the speed math. Composing aloud is a different skill from composing at a keyboard, because you cannot see the sentence you are building, and early dictated drafts tend to be wordier than typed ones. The advantage arrives as you learn to think a sentence ahead, which is why a fair trial of any dictation app lasts at least a week of real work rather than one demo paragraph. The study's participants were transcribing messages they were given, which removes the composing step entirely, so expect your own first numbers to fall short of theirs.
Dictation also loses outright in some situations, and it helps to know which before you blame an app. Editing existing text is slow by voice, because moving a cursor and selecting a word are easier with a mouse than with commands. Code is full of symbols and exact spellings. Very short replies cost more in key presses and waiting than they save. And shared rooms add a social cost and a technical one, because a microphone picks up the people around you. Where dictation wins clearly is a narrower but very common set of tasks:
- First drafts of emails, documents and reports, where flow matters more than precision
- Long messages in chat apps that you would otherwise shorten to save typing
- Notes and to-dos captured while you are looking at something else
- Prompts for AI tools, which reward long, specific, conversational instructions
The common thread is that each of these tasks is generative: you are producing new text from thought, not manipulating existing text. That is where the three-to-one speed ratio applies almost in full, and where the occasional wrong word costs one quick fix. The last item has become one of the most common uses of dictation in 2026, because a spoken prompt naturally carries the context and nuance that people trim when they type. Several apps in this guide now market themselves explicitly to people who write long prompts for AI assistants and coding tools, and the speed math explains why that audience converts so readily.
Why this matters: the value of a dictation app is decided by its error rate on your speech more than by any other single number, so a cheap, accurate app beats an expensive, feature-rich one that mishears you. How to apply it: time yourself once. Type a 150-word message you would really send, then dictate a similar one and fix it; if the dictated version, including corrections, is not clearly faster after a week of practice, the problem is accuracy, and a different engine is worth testing before you give up on dictation.
4. Where your voice goes: on-device or cloud
Your voice is more revealing than the text it becomes. A recording carries the sound of your voice, the room you are in, and everyone talking near you, and a dictation app hears everything you say while the key is held, including the sentence you abandon halfway and the colleague asking about lunch behind you. In our own accuracy tests (section 14), the recording that went most wrong was an 18-second meeting clip in which several other people were talking, and the recognizer faithfully transcribed their words too. Whatever the microphone hears goes wherever recognition runs.
That is why the most important question about a dictation app is not a feature but an architecture: is your speech recognized on your device, or uploaded to a vendor's servers? There is a useful distinction here between a policy and a property. A privacy policy is a promise: it can be well written and honestly meant, and it can still change, as Wispr Flow's data controls page notes that a setting "was previously called 'Privacy Mode'" and was last updated on August 18, 2026 - Wispr Flow. On-device recognition is a property you can verify yourself in one minute: turn off Wi-Fi and dictate.
Here is what each vendor says about where recognition happens and what is kept, in the vendors' own words or as close to them as a table allows:
| App | Where speech is recognized | What the vendor says about keeping it |
|---|---|---|
| Wispr Flow | "Transcription always occurs on the cloud" - Wispr Flow | Audio, transcripts and edits may be used to train models if you allow data sharing |
| Aqua Voice | In the cloud; "there is no offline mode" - Aqua Voice | May store transcripts when Privacy Mode is disabled - privacy policy |
| Willow Voice | Cloud, plus an offline mode "powered entirely by your device" - Willow | Claims zero data retention; enforced for teams on the Business plan |
| Windows voice typing | Online speech recognition "powered by Azure Speech services" - Microsoft Support | Not stored, sampled or listened to without your permission - Microsoft Support |
| Apple Dictation | On device where supported, otherwise Apple servers | Not stored unless you opt in to Improve Siri and Dictation - Apple privacy |
| MacWhisper | Local models on the Mac, or cloud models you choose - MacWhisper | Says local processing keeps sensitive content from ever leaving your Mac |
| VoiceInk | On the Mac - VoiceInk | Voice never leaves the Mac; optional cloud enhancement sends text only |
| Yap | On the Mac or PC | Nothing uploaded; history saved only on the device |
The table hides some nuance worth spelling out. Apple is a hybrid: its feature availability page lists 48 language and region entries for on-device Dictation out of 63 for Dictation overall, and for the others your words are sent to Apple's servers, though Apple says they are not stored unless you opt in. Microsoft draws the line by feature: voice typing is online, while the separate voice access feature "uses modern, on-device speech recognition" and "works even without the internet" - Microsoft Support. Microsoft's own screenshot of the voice typing settings shows the cloud connection plainly:
Look at the bottom of that panel: a link inviting you to contribute voice clips to help Microsoft "teach our speech technologies how people really talk", and the footer "Powered by Microsoft Online Speech Tech". Neither is sinister; contribution is optional, and Microsoft states that voice typing works without it. But the panel makes the architecture visible in a way most third-party apps do not: your speech is being processed by an online service, and the company running it has a standing interest in collecting samples to improve it.
Cloud recognition has real advantages, and a fair guide should say so. The most accurate English systems on the public Open ASR Leaderboard are cloud services: Zoom Scribe v2 Pro leads with a 3.59% average word error rate, ahead of Microsoft's Azure Speech at 3.81% - Hugging Face Open ASR Leaderboard. Cloud apps also run well on older hardware, which is why Superwhisper advises that Intel Macs "work best with Cloud models" - Superwhisper, and they can apply large language models for formatting that would not fit on a laptop. On-device recognition wins on everything that follows from never uploading audio: no transmission, no server retention to trust, recognition that works on a plane, and no per-minute server cost, which is one reason local apps can be cheaper.
Four questions settle where any app stands, and the vendor's own pages answer all of them if you look:
- Where is my speech recognized, and does that change by language or setting?
- Is anything stored on the vendor's servers, and for how long?
- Can my audio or text train models, and is that on by default?
- Does dictation still work with the internet turned off?
The fourth question is the one you can answer yourself, and it is the most reliable, because it tests behavior rather than wording. Be careful to check all three places your words can travel, not just recognition: an app can recognize speech locally and still send your text to a cloud language model for rewriting, or sync your dictation history to its servers for cross-device access, which Wispr Flow's page describes as a separate "Dictation Cloud Storage" control. We explain how on-device recognition works on a Mac in more depth in our guide to offline dictation for Mac, and our side-by-side comparison with Wispr Flow and comparison with Aqua Voice set the two cloud apps against Yap point by point, including where each of them is the better choice.
Why this matters: you will dictate things you would never type into a web form: client names, medical details, half-formed opinions, passwords spoken aloud by accident. Where that audio goes is the one property you cannot fix later without switching apps. How to apply it: decide first whether your work allows cloud processing at all; if it does not, your shortlist is the on-device column, and the airplane-mode test confirms it in a minute. If it does, prefer apps where model training is off unless you opt in, and read the data page, not the homepage.
5. Accuracy: what word error rate tells you, and what it hides
Accuracy in speech recognition is measured as word error rate, or WER: the number of substituted, deleted and inserted words, divided by the number of words actually spoken - Wikipedia. If you say "can we move the product review to Thursday afternoon" and the app writes "can we move the product view to Thursday afternoon", that is one substitution in nine words, an 11% error rate for that sentence. Lower is better, and at a 5% error rate you should expect a wrong word roughly every twenty words, which is one or two per paragraph.
A widely cited public comparison of recognition models is the Open ASR Leaderboard on Hugging Face, which averages each model's error rate across eight English test sets, from audiobooks and parliament speeches to meetings and earnings calls. We pulled its result data on October 5, 2026, from the revision updated that afternoon - leaderboard data on Hugging Face. The chart shows the two leading systems and ElevenLabs' Scribe, all proprietary cloud services, next to open models that local dictation apps can actually run.
Three things stand out. First, the gap is small: the best cloud system averages 3.59% and Parakeet TDT 0.6B v3, a model small enough to run on a laptop, averages 4.86%, a difference of about one word in eighty. Second, the best open model, Qwen3-ASR 1.7B, now sits within a percentage point of the leaders. Third, Whisper large-v3, still the engine behind many local apps, trails the newer models at 5.78%, and the faster turbo variant gives up a little more. If an app offers you a choice of models, those differences are worth a setting change.
The leaderboard average also hides how much the type of speech matters. The same data shows Parakeet v3 at 3.13% on the harder LibriSpeech "other" audiobook set but 9.42% on AMI meeting recordings, and Whisper large-v3 at 3.52% against 13.63%. Conversation, with its overlaps, false starts and distant microphones, is several times harder than one person reading clearly. Dictation sits closer to reading than to meetings, since you are one speaker close to the microphone, but the lesson holds: an average score tells you about the model, not about your room.
We ran our own measurement when we chose the engines for Yap, on 60 public recordings selected before any model ran: 20 English audiobook clips from LibriSpeech's harder test set, 20 Dutch clips from Google's FLEURS, and 20 English meeting clips from the AMI corpus. The chart compares the Whisper Base model Yap used before the upgrade with the two Parakeet engines it ships today: Parakeet Ultra on Mac, a post-trained version of Parakeet v3 from Moondream - Moondream on Hugging Face, and the original Parakeet v3, in an 8-bit version, on Windows.
The results are not uniformly flattering, which is why we show them. Both Parakeet engines cut errors sharply against the old engine, most dramatically in Dutch, where Whisper Base got a third of the words wrong. But on meetings, the Mac engine scored worse than the Windows engine, and the cause is a single 18-second clip in which several people talk at once: the Mac engine transcribed the background speakers too, producing 32 of its 78 errors on that one clip. We kept it in the total because real rooms have other people in them. Our samples are small, about eight minutes of audio in all, so read them as a sanity check on the public numbers rather than a ranking.
The leaderboard and our tests share two blind spots you should know about. Both ignore punctuation and capitalization when scoring, so formatting quality, which you will notice in every sentence, is invisible to them; NVIDIA's own model card notes that its accuracy figure "does not include Punctuation and Capitalisation errors". And neither covers your vocabulary. Recognition errors cluster on names, product terms and jargon, which is why several apps add custom dictionaries: Wispr Flow "learns your names and jargon", VoiceInk lets you add custom words, and Aqua Voice says its client-side Deep Context engine "reads what's on screen" to get names and technical terms right - Aqua Voice. Yap does not offer a custom dictionary today, which is a real gap if your work is full of unusual names.
The only accuracy number that matters is the one on your voice, and you can measure it in about ten minutes per app:
- Write a 150-word script with your own names, jargon, numbers and one long sentence.
- Read it twice into each app: once in a quiet room, once with a podcast playing.
- Count the errors by comparing against the script, word by word.
- Dictate for 40 seconds without a script, the way you really compose.
- Repeat in each language you write in, and in the apps you use most.
The script makes apps comparable, and the unscripted ramble shows how they handle your real pace, pauses and restarts. Divide the error count by the number of words to get your personal word error rate, then apply the break-even rule from section 3: under about 5% you will love dictating, above about 10% you probably will not. Pay special attention to errors on names and numbers, because those are the mistakes that are both most likely and most embarrassing to send. If two apps tie on your script, the noisy-room run usually separates them.
Why this matters: published accuracy numbers describe models on someone else's recordings, while your experience depends on your accent, microphone, room and vocabulary, and those can easily double or halve an error rate. How to apply it: shortlist apps by engine quality using the numbers above, then decide with your own ten-minute test; never pay for an annual plan before you have counted the errors on your own voice.
6. Speed: the wait after you stop talking
There are two kinds of speed in dictation, and only one of them depends on the app. Input speed is how fast you talk, and it is the same in every app. Turnaround is the time between the moment you stop talking and the moment finished text appears where your cursor is. In push-to-talk dictation the app waits for you to finish, then does its work, so turnaround is a pause you feel at the end of every single thought. It is the number that decides whether dictation feels like typing or like waiting.
Usability research gives a useful yardstick. Jakob Nielsen's classic response-time limits put 0.1 seconds as the limit for feeling that a system reacts instantly, and 1.0 second as "about the limit for the user's flow of thought to stay uninterrupted, even though the user will notice the delay" - Nielsen Norman Group. No dictation app reaches the first limit, because recognition takes real computation. The practical goal is the second: text that lands in under about a second keeps you in your train of thought, while two or three seconds turns every sentence into a small wait you start to notice.
The wait is made of five parts, and each one can dominate depending on the app and the situation:
- End of speech: the app has to register that you are done, from a key release or detected silence
- Transport: cloud apps upload your audio and download the text
- Recognition: the model runs on a server or on your own chip
- Formatting: rules are nearly instant; a language model pass adds time
- Insertion: the text is typed or pasted into the app in front
For a cloud app, transport is the wild card. Ten seconds of speech at 16 kHz in 16-bit mono is about 320 kilobytes before any compression, which a good connection moves in a fraction of a second, but hotel Wi-Fi, a phone hotspot or a congested office network can stretch the round trip unpredictably. For an on-device app there is no transport, and the variable becomes your hardware: the same recognition work takes longer on an older or less powerful chip. Formatting with a large language model adds a second inference step on top of recognition, which is why apps that rewrite your words have to engineer hard to stay under a second.
Vendors do publish speed figures, but they are targets and best cases rather than measurements you can compare. Wispr Flow's engineering team writes that its users expect "full transcription and LLM formatting/interpretation of their speech within 700ms of when they stop speaking" - Wispr Flow. Aqua Voice says its Instant Mode puts text on screen about 450 milliseconds after you stop speaking, and that it inserts text "in about a second (as fast as ~450ms)" - Aqua Voice. We measured Yap on two Macs, from releasing the shortcut until the finished text was ready, over a mix of English, Dutch, a 40-second recording and a dictation right after the speech models had been unloaded.
On an Apple M5 Pro, Yap's median was 0.56 seconds, with individual dictations between 0.28 and 0.93 seconds. On an Apple M2 Pro, the same version took a median of 1.44 seconds, between 1.23 and 2.66 seconds, about two and a half times longer. That spread is the honest cost of on-device recognition: your speed is your chip's speed. A cloud app's speed, by contrast, is mostly its servers' and your network's, which is why the same cloud app can feel instant at the office and sluggish on a train. Neither model is faster in general; they are fast in different conditions.
Cold starts are the hidden variable in on-device speed. A local app has to load its speech model into memory before it can recognize anything, and to save memory many apps unload the model after a period of inactivity. Yap, for instance, releases the model's memory after a minute without dictation and starts preparing it again the moment you start recording, so the loading overlaps with your speaking. The first dictation after a long break is the one to time if you want to know the worst case, and our sample deliberately included one.
Recording length matters less than you might expect on a fast chip. Our sample mixed short sentences with a 40-second recording, and all seven dictations finished within 0.93 seconds on the M5 Pro. On a cloud app, a longer recording also means a longer upload, so the same test on a weak connection is worth running before you commit to an app for long-form drafting.
On Windows the variable is the processor rather than a dedicated neural engine. Yap's Windows recognizer runs on the CPU with at most four threads, deliberately leaving a core free for the app you are typing into, so a modern laptop processor matters more than a graphics card. That design keeps the computer responsive while it works, at the cost of a slightly longer wait on older machines, which is the kind of trade-off worth asking any on-device vendor about.
Some apps hide the wait by showing words while you talk. Aqua Voice's Realtime Mode, for example, makes "words appear in real time as you talk" before the final text settles. Live text is reassuring, but the finished, formatted version still arrives after you stop, and the words can change in between, which some people find distracting when they glance at the screen mid-sentence. Whether you prefer live words or a clean result after a short pause is a matter of taste worth testing, not a measure of which app is faster.
Why this matters: a turnaround over about a second breaks your flow on every sentence, and over a working day that friction decides whether you keep dictating. How to apply it: measure it yourself rather than trusting any chart, including ours. Film your screen with a phone in slow motion, release the key, and count the frames until the text appears; repeat ten times, on battery, and, for a cloud app, once on a weak network. If the slowest of your ten tries is still under two seconds, speed will not be your problem.
7. Languages, accents and switching between them
Language support is where vendor numbers are least comparable, because vendors count different things. Some count languages, some count language and region entries such as English (India) and English (United Kingdom) separately, and model cards count the languages a model was trained on, which an app may or may not expose. A larger number is not automatically better either: what matters is whether your languages are covered well, whether the app notices when you switch, and whether it leaves your words in the language you spoke.
Here is how each option states its coverage, with the caveat that the units differ from row to row. Region entries matter if you speak a regional variety of a language: a vendor that lists English (India) or Portuguese (Brazil) separately is telling you that it supports that variety explicitly, which a bare language count cannot.
| App or model | Coverage as the vendor states it |
|---|---|
| Wispr Flow | 100+ languages, detected automatically |
| Willow Voice | Over 100 languages, switch without missing a beat |
| Superwhisper | 100+ languages and dialects, with translation to English |
| Apple Dictation | 63 language and region entries, 48 of them on device |
| Aqua Voice | 49 dictation languages, with an Auto setting |
| Windows voice typing | 46 language and region entries |
| Parakeet v3 and Yap | 25 European languages, detected per recording |
A count says nothing about quality per language, and the gap between English and everything else is larger than most people expect. NVIDIA's own chart for Parakeet v3 shows average error rates of roughly 11 to 12 percent across 24 languages on the FLEURS benchmark, for both Parakeet and Whisper large-v3, more than twice what the same models average in English:
The chart also shows that size still buys accuracy: the 2.3-billion-parameter SeamlessM4T v2 model beats both Parakeet and Whisper on all three benchmarks, at nearly four times Parakeet's size. Averages also hide the spread between languages, since languages with plenty of training data can fare much better than those with little. Post-training helps: Moondream reports that its Parakeet Ultra lowers the 25-language FLEURS error rate from 11.62% for the original model to 9.55% - Moondream. In our own Dutch test, the Mac engine got 4.38% of words wrong, far better than the multilingual average, which is a reminder that your language may be well above or well below the mean.
How an app decides which language you are speaking matters as much as the list. There are three approaches. With a language picker, you choose in settings, which is reliable but means switching by hand: on a Mac that happens in Keyboard settings or with the Globe key, as our guide to changing the dictation language on a Mac explains, and on Windows you change the input language. With automatic detection per recording, the app listens to each recording and decides; Parakeet v3 does this by itself, and Wispr Flow "automatically detects and transcribes the language you're speaking" - Wispr Flow. And mixing languages inside one sentence remains unreliable in general, because a short or ambiguous phrase gives the model little to decide on.
Watch for translation too. Superwhisper advertises that it can translate 100+ languages to English, which is exactly right if you want English output and exactly wrong if you write to a Dutch client and expect Dutch back. Yap transcribes every recording in the language it was spoken in and never translates, which is the behavior we wanted for people who switch between languages all day, but it is a choice with a cost: if you speak a language outside its 25, it is the wrong app, and a Whisper-based or cloud app with broader coverage is the better pick.
Accents are the least documented part of this whole market. None of the apps in this guide publishes accuracy by accent, and the public benchmarks only partly reflect the range of real speakers. The ten-minute test in section 5 is the only reliable answer, and it is worth running with the specific words your accent tends to trip on, such as names, numbers and words borrowed from your first language.
Why this matters: if you write in more than one language, a wrong guess about your language produces gibberish rather than a typo, and an app that forces you into settings every time you switch will quietly stop being used. How to apply it: list every language you write in, confirm each one appears on the vendor's list with your region, then dictate one sentence in each, back to back, without touching any settings, and check that every sentence comes out in the language you spoke.
8. Formatting, voice commands and AI rewriting
Raw recognition produces words. Formatting is everything that turns them into text you would send: punctuation, capital letters, numbers as digits, paragraph breaks, and the removal of verbal clutter. There are three ways apps do it, often combined. The recognition model can produce punctuation itself, as Parakeet v3 does. The app can listen for spoken commands, such as saying "comma" or "new paragraph". Or a language model can rewrite the transcript into cleaner prose. Each approach has a different failure mode, and choosing between them is mostly a question of whether you want exactly what you said or what you meant.
Spoken commands have a problem that is easy to miss: command words are also ordinary words. "The trial period ends on Friday" should keep the word, not insert a full stop, and a model has to guess which you meant. Microsoft's voice typing supports commands such as "delete that" and "select the previous word" - Microsoft Support, and Word's Dictate feature lists punctuation phrases for each language it supports. Yap takes the opposite default, as its Windows settings show:
With spoken punctuation off, words like "period", "colon" and "new paragraph" stay words, and punctuation comes from the recognition model; turning it on makes them English formatting commands. The screen also shows automatic language detection, which has no setting to change because there is nothing to pick, and the hold-to-record shortcut, Ctrl and the Windows key held together. These are product choices rather than universal truths: if you dictate structured documents full of lists and headings, an app with a rich command vocabulary may suit you better.
AI rewriting is the fastest-moving part of the market. Wispr Flow builds language-model formatting into every dictation and designs for the full result within 700 milliseconds. Superwhisper offers modes that "optimize tone, structure, and formatting", and its Pro plan lets you use your own AI API keys. Willow Voice sells a writing feature called Scribe, and Aqua Voice describes its output as speech recognition "fused with language-model post-processing" - Aqua Voice. Microsoft has gone local with fluid dictation on Copilot+ PCs, which corrects grammar, punctuation and filler words using "on-device small language models (SLMs)", in all English locales - Microsoft Support.
Rewriting has three costs worth weighing from first principles. A rewrite model optimizes for plausible text, not for your exact words, so it can drop a hedge, shift a tone, or turn an unusual name into a common word, which is harmless in a chat message and risky in a quote, a contract note or a medical record. It usually runs on a cloud language model, so an app can recognize your speech locally and still send your text to a server for cleanup; VoiceInk is explicit that its optional cloud enhancement sends "only transcribed text (not your voice)" - VoiceInk. And it adds a second inference step, which costs time unless the vendor has engineered around it.
If rewriting is the feature you want and the cloud is the part you do not, the alternative is a language model running on your own computer, and Founden's scored comparison of private and local AI options covers desktop apps such as Jan and LM Studio, which run open-weight models on your own machine. Founden is run by the same founder as Yap, and this site is built on it.
Why this matters: formatting is the part of dictation you see in every sentence, and the wrong default, whether commands that eat your words or a rewriter that changes them, creates corrections that cancel the speed advantage. How to apply it: decide per context. For notes, quotes and anything legal or medical, use verbatim output and check that rewriting can be switched off; for messages and emails, try a rewriting mode for a week and keep it only if you rarely have to undo what it changed. In both cases, find out where the rewriting runs before you dictate anything confidential.
9. Typing into any app, safely
There are two kinds of dictation tools, and the difference shows up the moment you leave a single document. In-app dictation works inside one program: Word's Dictate needs "a microphone and reliable internet connection" and an active Microsoft 365 subscription - Microsoft Support, while Google Docs voice typing works in the latest versions of Chrome, Edge and Safari, where "your web browser controls the speech-to-text service" - Google Docs Help. System-wide dictation types into whatever app is in front: your email, a chat window, a form in a browser, a terminal.
System-wide dictation is far more useful and far harder to do safely, because the app has to decide where the text goes at the end of a recording you started several seconds earlier. Three insertion methods exist. Simulated keystrokes type the text character by character, which works almost everywhere but can collide with apps that treat key presses as commands. Clipboard pasting is fast for long text but replaces whatever you had copied, unless the app restores it afterwards. And accessibility insertion sets the text directly in fields that support it. Each method fails differently, and good apps combine them with checks.
The checks are what separate a safe app from a risky one. On Windows, Yap's insertion helper is tested to refuse three situations: a password field, an open menu, and a window that is no longer the one you were in when you started speaking. In each case, the text is held back rather than typed somewhere you did not intend. Every dictation is also saved to a local history before insertion is attempted, so a refusal or a failure never loses your words:
Each entry in that list carries its outcome, "Sent to editor" here, and a copy button for the cases where the text could not be typed. A history is the cheapest insurance a dictation app can offer, and it is worth checking where it lives: Yap keeps it only on the device, while Wispr Flow offers an optional "Dictation Cloud Storage" for cross-device access. Neither is wrong, but they answer the privacy question in section 4 differently.
The shortcut is the other half of insertion, because it decides how you start and stop. Apple Dictation starts with the Microphone key, a shortcut you choose, or Edit > Start Dictation. Windows voice typing opens with Win+H. Superwhisper starts with Option and Space. Yap uses a hold-to-talk shortcut, Fn on a Mac and Ctrl+Win on Windows, with Space to lock recording for hands-free dictation. Modifier-key shortcuts are harder to build than they look: on Windows, releasing the Windows key by itself opens the Start menu, and releasing Alt can open an app's menu bar, so an app that uses them must suppress those side effects, which Yap's Windows test suite checks.
Undo is a small test with a large effect on daily use. In many editors, text pasted in one operation can be undone in one step, while text typed character by character may take several. If you often glance at a dictation and decide to say it again, check how many presses of Undo it takes to remove one in your main editor, because that number is how much a bad result costs you every time.
On a Mac, a system-wide dictation app asks for two permissions: the microphone, and Accessibility to type into other apps. Both are reasonable requests for this kind of tool, but Accessibility is a powerful permission, since it lets an app control other apps, which is one more reason to prefer developers who explain exactly what their app does with it and when.
Chat boxes and terminals deserve their own check, especially if you dictate prompts for AI tools. Many chat inputs send the message when they receive an Enter key, so a dictation that contains a line break can submit half a thought, and some terminals treat pasted and typed text differently. If those are the places you dictate most, test multi-line dictation there before anything else.
A five-minute insertion test catches nearly every problem before it costs you:
- Dictate into your five most-used apps, including a browser form and a chat window.
- Try a password field and a terminal, and see what the app does.
- Click into another window halfway through a sentence.
- Copy something first, dictate, then paste, to see whether your clipboard survived.
- Dictate a long paragraph and look for dropped, doubled or reordered characters.
The third and fourth checks are the ones most apps fail quietly. An app that types into whichever window is in front at the end of the recording will happily put half an email into a team chat if you clicked there while thinking. An app that pastes through the clipboard without restoring it will replace the link you copied five minutes ago, and you will only notice when you paste. Neither failure is visible in a demo, and both are the kind of thing that makes people stop trusting dictation without being able to say why.
Why this matters: insertion is the last step, and a failure there either loses your words or puts them somewhere you did not intend, which is worse than a typo because it can send private text to the wrong place. How to apply it: run the five-minute test on every finalist, prefer apps that keep a history you can recover from, and treat any app that types into a password field or loses your clipboard as disqualified, whatever its accuracy.
10. What dictation apps really cost
Dictation apps are sold in three ways: as a subscription, as a one-time purchase, or for free, either built into the operating system or as a capped free tier. The pricing model is not arbitrary, and it follows from the architecture in section 4. A cloud app pays for server computation every time you speak, so it has a running cost per user and almost has to charge a subscription. An on-device app has no per-use cost, because your own chip does the work, which is what makes one-time pricing possible at all.
That does not make every on-device app cheap, and we should be direct about our own case. Yap recognizes speech on your device and is still sold by subscription, at $6.90 a month or $69 a year after 3 free days with a free account and no card, and a 10-day trial that requires a card - Yap pricing. The subscription pays for continued development and updates rather than servers. If you would rather pay once for on-device dictation on a Mac, VoiceInk is the app to look at. Here is what each option costs as of October 5, 2026, from the vendors' own pricing pages:
| App | Free option | Paid monthly | Paid yearly | One-time |
|---|---|---|---|---|
| Wispr Flow | 2,000 words a week on desktop | $15 | $12 a month ($144) | - |
| Willow Voice | Unlimited, on a weaker model | $15 | $12 a month ($144) | - |
| Aqua Voice | 1,000 words | $10 | $8 a month ($96) | - |
| Superwhisper | Unlimited Whisper models | $8.49 | Two months free | Lifetime license offered |
| Yap | 3 days free, then a 10-day trial | $6.90 | $69 | - |
| VoiceInk | Free trial | - | - | $25 one Mac, $39 two, $49 three |
| MacWhisper | Free version | - | - | €64 |
| Apple Dictation, Windows voice typing | Free, built in | - | - | - |
Superwhisper's yearly and lifetime prices are shown only in an interactive toggle on its page, so we list what its page text states rather than guess. Dragon Professional's product page shows no price at all, only a "Contact us" button - Nuance, and Word's Dictate is included with a Microsoft 365 subscription rather than sold on its own. For the four subscription apps with both monthly and yearly prices, the chart shows what twelve months cost each way:
Free tiers deserve a closer look, because their limits are measured in words and words go faster by voice than people expect. At a relaxed 150 words per minute, Wispr Flow's free 2,000 words a week is about 13 minutes of speaking, and Aqua Voice's 1,000 free words is about seven minutes in total. Those are generous trials and thin plans for daily use. Superwhisper's free tier is different in kind: unlimited dictation on local Whisper models, with paid features layered on top, which is why it scores so well on cost. Willow Voice also offers unlimited free dictation, on what it calls a weaker model.
Over several years the gap between subscriptions and one-time purchases widens steadily, which is the comparison a line chart shows best:
Price still matters less than fit, and the time math from section 3 shows why. Suppose you produce 1,000 words a day of text you could dictate. Typing it at 52 words per minute takes about 19 minutes; dictating it at 153 words per minute takes about 6.5 minutes, plus about 4 minutes of fixes at a 5% error rate. That saves roughly 8.5 minutes a day, or about 35 hours over 250 working days. Against that, even the most expensive plan here costs about $4 per hour saved, and Yap about $2. The expensive mistake is not the subscription price; it is an app whose error rate on your voice erases the saving, or whose data handling your work does not allow.
One-time purchases carry a different risk: an app that is never paid for again has to fund its updates from new sales, and operating system updates do break apps. VoiceInk's answer is that it is built in public, with its code on GitHub - VoiceInk on GitHub. Subscriptions carry the opposite risk: prices can rise, and features can move between tiers. Neither model is safe from change; they fail in different directions.
Why this matters: the sticker price is the most visible number in this market and one of the least important, unless your budget is tight, in which case the free and one-time options in this guide are genuinely good. How to apply it: compare plans on your real usage, count how many words a week you would dictate before trusting a free tier, prefer yearly billing only after a full trial month, and weigh the cost of the wrong app in lost time, not just the price of the right one.
11. Mac, Windows or phone: platform coverage
Platform support is a hard filter rather than a preference: an app that does not run on your computer scores zero for you, whatever the table says. The question has three layers. Does it run on both of the operating systems you use? Does it support your specific hardware, especially Intel Macs and Windows on Arm laptops? And do you need the same dictation, with the same vocabulary and formatting, on your phone? Each layer removes a different set of apps.
The hardware layer catches people out more than the operating system does. On-device recognition needs a capable chip, so several local apps simply do not support older hardware: VoiceInk "requires Apple Silicon and macOS 14.4 or later", Yap runs on Apple Silicon Macs, and Superwhisper's offline models "only run really well on Apple Silicon macs". Cloud apps are more forgiving because your computer only records and uploads; Aqua Voice, for example, supports both Apple Silicon and Intel Macs. On Windows, Yap ships separate installers for standard x64 PCs and Arm-based laptops, and its speech model downloads once, about 0.7 GB.
| App | Mac | Windows | iPhone | Android |
|---|---|---|---|---|
| Wispr Flow | Yes | Yes | Yes | Yes |
| Superwhisper | Yes; offline models best on Apple Silicon | Windows 10 and 11 | Yes | Yes |
| Aqua Voice | Apple Silicon and Intel | Windows 10 and 11 | Yes | Waitlist |
| Willow Voice | Yes | Yes | Yes | Not listed |
| Yap | Apple Silicon | Windows 10 and 11, x64 and Arm | No | No |
| VoiceInk | Apple Silicon, macOS 14.4 or later | No | New iOS app | No |
| MacWhisper | Yes | Not listed | Not listed | Not listed |
| Dragon Professional | Discontinued in 2018 | Windows 10 and 11 | Dragon Anywhere app | Dragon Anywhere app |
Dragon deserves a note because it is still the name many people search for. Nuance discontinued Dragon Professional Individual for Mac effective October 22, 2018, and stopped providing updates after that date - Nuance. Dragon Professional 16 remains a Windows product, "optimized for Windows 11 and backwards-compatible to Windows 10" - Nuance. Nuance's separate mobile app, Dragon Anywhere, runs on iOS and Android - Dragon Anywhere. We cover the Mac options in detail in Dragon for Mac: discontinued, and what to use now.
Phones are where cloud apps have a structural advantage. One account can carry your custom vocabulary, snippets and style preferences from your laptop to your phone, which is why Wispr Flow and Superwhisper both ship on all four platforms. A local app has to ship and run its model separately on every device, and phones have tighter limits on memory and battery than laptops. Before you pay for phone coverage, though, try the dictation built into your phone's keyboard: if it is good enough for short messages, a desktop-only app may cover everything else you write.
Why this matters: platform gaps are the most common reason people end up running two dictation apps with two different habits, two vocabularies and two privacy models. How to apply it: list every device you write on, including the work laptop you do not control, and remove any app that does not cover the ones you use daily; if you use both a Mac and a Windows PC, favor an app that behaves the same way on both, which is the problem Yap for Windows was built to solve.
12. When the built-in tools are enough
For many people, the right dictation app is the one already installed. Apple Dictation is free, types into any app on a Mac, and runs on device for 48 of its 63 language and region entries. You start it with the Microphone key, a shortcut you choose, or Edit > Start Dictation; you can dictate text of any length without a timeout, and ambiguous words are underlined in blue so you can pick an alternative - Apple Support. Auto-punctuation can be turned off in Keyboard settings. Our step-by-step guide to voice to text on a Mac covers setup, and our page on Mac dictation not working collects Apple's fixes by symptom.
macOS 27, released on September 14, 2026, made Apple Dictation better on some Macs. On a Mac with an M3 chip or later and at least 12 GB of memory, its more advanced Dictation "captures what users say as polished text with greater precision, automatically handling capitalization, punctuation, and formatting as they speak", in English only so far - Apple Newsroom. Every M1 and M2 Mac keeps the standard engine, and so does every other language. Our page on macOS 27 dictation shows which Macs get it.
Whether Apple Dictation is private for you depends on your language, and Apple tells you directly. Its privacy page says that your device "will indicate in Keyboard Settings if your audio and transcripts are processed on your device and not sent to Apple servers" - Apple privacy. So the check takes a minute: open the Dictation section of Keyboard settings with your language selected and read the note. If your language is processed on device, Apple Dictation already passes the on-device test that no cloud app can pass.
Windows voice typing is just as close at hand: press Win+H in any text field. It uses Microsoft's online recognition, so it needs an internet connection, and it supports 46 language and region entries, automatic punctuation, a profanity filter and commands such as "delete that". On Copilot+ PCs, fluid dictation adds on-device cleanup of grammar, punctuation and filler words in English, which Microsoft says reduces the need for manual editing.
Windows also has voice access, which is built mainly for controlling a PC by voice but can author text too. Microsoft says it "uses modern, on-device speech recognition" and "works even without the internet", and it is available from Windows 11 version 22H2, where it replaces the older Windows Speech Recognition - Microsoft Support. For someone who needs offline dictation on Windows occasionally and does not want to install anything, it is the first thing to try, with the caveat that it is built first for controlling the PC rather than for fast drafting.
Inside documents, Word's Dictate and Google Docs voice typing work well if most of what you write lives in those two places. Microsoft states that Word's dictation service "does not store your audio data or transcribed text", though it needs an internet connection and a Microsoft 365 subscription. Google Docs voice typing leaves the speech processing to your browser, as section 9 explained, and works in the latest Chrome, Edge and Safari. Neither helps with email in another app, chat or forms, which is where system-wide tools earn their keep.
The built-in tools have real limits, though, and three of them account for most of the reasons people move on:
- You need it offline on Windows for real drafting, where voice typing requires the internet
- Your language is not on device on your Mac, so your words go to Apple's servers
- Your error rate is high on the ten-minute test in section 5
Any one of these is a good reason to try an app. Two quieter limits often come along with them: switching languages means changing settings or input languages by hand, and neither built-in tool keeps a history you can recover a lost dictation from. The error rate deserves the most attention, because people tend to blame themselves for mumbling when a different engine would recognize them fine. The reverse is also true, and worth saying in a guide written by a vendor: if Apple Dictation runs on device in your language, recognizes you well and you mostly dictate short messages, there is no reason to pay anyone, including us.
Even if you end up paying for an app, the built-in tool is the right baseline for your tests. Run the ten-minute script through it first and write down its error count; any app you pay for should beat that number clearly, on the same script and the same microphone, or it is not worth the money. A baseline also protects you against a common trap in software buying, which is comparing a new app's best moment with the old tool's worst one.
Why this matters: paying for a problem you do not have is the most common way people overspend on software, and dictation is no exception. How to apply it: use the built-in tool for a week first. Write down every moment it fails you, then use that list as your test plan for paid apps; if the list stays empty, you are done.
13. A decision framework
Most of the decision comes down to four questions asked in order, because each one removes options that no later answer can bring back. Privacy comes first because it is a constraint of your work rather than a preference: if your words cannot leave your computer, half the market is out before features enter the conversation. Phone coverage comes next because it splits the remaining apps cleanly, then budget, and only then the comparisons of accuracy and speed, which you settle by testing rather than by reading.
The diagram turns those questions into a shortlist. It does not pick a winner for you, because the winner depends on your voice and your rooms, which no chart can see. Treat each end point as the two or three apps to put through the ten-minute accuracy test and the five-minute insertion test before you pay.
A few common situations show how the framework plays out. A lawyer, clinician or consultant under confidentiality should start from the on-device branch and verify it with the airplane-mode test, then prefer verbatim output over rewriting; on a Mac alone, VoiceInk is the cheapest way in, and if the same person also uses a Windows PC, Yap gives them one app that behaves the same on both. A writer on two devices, drafting on a laptop and replying on a phone, should start with Wispr Flow, which reaches the most platforms, and Superwhisper, which covers the same four and adds offline models on the desktop.
A multilingual writer should check language coverage before anything else. If every language they write is among the 25 that Parakeet covers, local engines do very well, as our Dutch results show; if not, a Whisper-based or cloud app with broader coverage is the safer choice, and that rules Yap out. Anyone on a tight budget, students included, has three genuinely free routes: the built-in tools, Superwhisper's free tier on local Whisper models, and Willow Voice's free plan on its weaker model. Two vendors also discount for students at the time of writing: Superwhisper advertises 40% off its Pro plan for students, and Aqua Voice says students get 70% off Pro and Max - Aqua Voice. The free routes are worth exhausting first, because the ten-minute test will tell you quickly whether a paid engine recognizes you meaningfully better.
Someone who mostly writes long prompts for AI tools will get the most from apps with good formatting and context features, which today are mostly the cloud apps, with the privacy trade-off that implies. And someone who dictates a few messages a day should simply start with the built-in tool.
Why this matters: a structured decision prevents the most common failure in software buying, which is choosing on the most visible feature and discovering the deal-breaker a month later. How to apply it: answer the four questions honestly, take the two or three apps at your end point, and spend thirty minutes running all of them through the tests in sections 5 and 9 on the same day, with the same script and the same microphone.
14. How we researched and tested this guide
Every vendor fact in this guide, from prices to privacy policies to platform lists, was read on the vendor's own pages on October 5, 2026, and each one links to the page it came from. Where a sentence carries a privacy claim, we quote the vendor rather than paraphrase it, because the exact wording is usually where the meaning lives. Public benchmark figures come from the Open ASR Leaderboard's result data as revised that afternoon, NVIDIA's model card for Parakeet TDT 0.6B v3, and Moondream's card for Parakeet Ultra. We did not include apps whose pricing or data handling we could not verify from primary sources.
Our accuracy figures come from 60 public recordings, selected before any model ran: 20 evenly spaced clips from LibriSpeech's harder "other" test set - LibriSpeech, 20 from the Dutch test set of Google's FLEURS - FLEURS, and 20 English meeting clips from the AMI corpus between 4 and 30 seconds long - AMI. Together they make about eight minutes of human speech. Scoring normalizes letter case and ignores punctuation, the same way for every engine. No recording was removed and no reference was rewritten, including the meeting clip with background speakers that hurt our Mac engine's score.
The engines were measured in the runtimes Yap ships. Parakeet Ultra ran through Apple's Core ML using the FluidAudio runtime - FluidAudio on GitHub. The Windows engine, Parakeet TDT 0.6B v3 in its 8-bit version, ran through sherpa-onnx - sherpa-onnx on GitHub with the same model files and runtime the Windows app uses, executed on a Mac's processor. Speed was measured by the installed app, version 0.7.36, from releasing the shortcut until the finished text was ready, with prerecorded speech passing through the app's real recording, recognition and formatting path on an Apple M5 Pro and an Apple M2 Pro.
There are things we did not test, and you should know them. We did not run competitor apps on our recordings: their cloud services, accounts and settings make an identical automated test hard to do fairly, and a half-fair comparison would be worse than none, which is why the scorecard scores engines on public evidence instead. We did not test real microphones in noisy rooms, accents beyond those in the public recordings, or speed across a range of consumer Windows PCs. Our samples are small, and the numbers are a check on the public benchmarks rather than a substitute for them.
Finally, the conflict of interest, stated once more where it belongs. We make Yap, and we wrote this guide. We scored Yap with the same criteria and the same standard of evidence as every other app, and it placed fourth; the cells explain why. If a vendor changes a price, a plan or a policy, or you find a mistake, tell us through the contact options on this site and we will correct the guide and its date.
Why this matters: a buyer's guide is only as useful as your ability to check it, and the method is what separates evidence from marketing. How to apply it: use the links to verify anything that would change your decision, rerun our accuracy test on your own recordings if you want numbers for your voice, and discount any claim, ours included, that you cannot trace back to a source.
Conclusion
Choosing a dictation app is easier once you see that the decisions come in a fixed order. First, decide where your speech may be processed, because that single architectural choice determines privacy, offline use and much of the pricing. Second, measure accuracy on your own voice, because the error rate decides whether dictation saves you time at all, and published numbers describe other people's recordings. Third, check reach: the platforms you write on and the apps you type into. Speed and price matter, but in most cases they settle themselves once the first three are right.
On our scorecard, Superwhisper and VoiceInk come out on top by pairing local recognition with low prices, and the free Apple Dictation is a strong third that many Mac users never need to replace. Yap places fourth: the most private option in the table and the one that behaves the same on a Mac and a Windows PC, held back by its price and the lack of a phone app. Wispr Flow is the most widely available and the right starting point if your phone matters as much as your laptop, and Aqua Voice and Willow Voice are worth testing if cloud processing is acceptable to you and you want polished, AI-formatted output.
The practical next step takes about an hour. Run the ten-minute accuracy test from section 5 on your top two or three candidates, the five-minute insertion test from section 9, and the airplane-mode test from section 4 if privacy matters to you. Keep the app that wins on your voice, not ours. If on-device dictation that works the same way on a Mac and a Windows PC is what you need, Yap works free for 3 days with a free account and no card, from the download page; if you only use a Mac and want to pay once, try VoiceInk first; and if you need your phone too, start with Wispr Flow or Superwhisper.
This guide reflects dictation apps, prices, policies and benchmark data as of October 5, 2026. Vendors change plans, prices and data practices often, so check the linked pages before you buy.