AI Dictation Accuracy 2026: What Affects It and How to Test

Why Accuracy Claims Are Hard to Compare#

Every dictation app says it's accurate. Few say how they measured it, and the ones that publish numbers rarely measure the same thing.

An error rate only means something next to the recordings it was measured on. Clean audiobook narration produces low numbers. A laptop microphone in a kitchen, a strong accent or a sentence full of product names produces much higher ones. Two apps tested on different recordings can't be ranked against each other, even when both numbers are honest.

So this guide doesn't publish a league table. It explains what actually drives dictation accuracy, what the published research says, and how to run a fair comparison with your own voice, which is the only test that predicts how an app will work for you.

Disclosure: LumeVoice publishes this blog and is one of the apps discussed. The comparison below sticks to what each company documents, with links so you can check it.

How Accuracy Is Measured#

Word error rate#

Speech recognition researchers measure accuracy with word error rate: compare the transcript with what was actually said, count the mistakes, and divide by the number of words spoken.

Word error rate = (substituted words + missing words + extra words) ÷ words actually spoken

For example, if you said 100 words and the transcript has two wrong words, one missing word and one extra word, the word error rate is 4 in 100, or 4%.

Most benchmarks ignore capitalisation and punctuation, and many normalise spelling (so "colour" and "color" match) before counting. Those choices alone can move a result. OpenAI's Whisper paper describes a dedicated text normaliser for exactly this reason, and it's one more reason numbers from different sources don't line up.

What benchmarks measure, and what they don't#

Many published figures come from public test sets. LibriSpeech, one of the best known, is made of audiobook recordings: clear voices, careful pacing, quiet rooms. Real dictation is messier. You think while you talk, restart sentences, say "um" and use words no audiobook contains.

There's a second gap that matters for dictation apps in particular. Word error rate rewards a transcript for matching your speech exactly, filler words and false starts included. Many dictation apps now deliberately remove "um", drop the half of a sentence you took back, and fix grammar. Against a word-for-word reference, that cleanup counts as errors, even though it's what you wanted. For dictation, the more useful question is how much editing the text needs before you can send it.

What Affects Dictation Accuracy#

The speech model#

The model that turns audio into text sets the ceiling. OpenAI's Whisper models, released as open source in 2022, sit behind many current apps; others add models such as NVIDIA's Parakeet or Mistral's Voxtral. Within a family, larger models are generally more accurate, and slower and more memory-hungry when they run on your own computer.

Your microphone#

A laptop's built-in microphone sits next to the keyboard and fans, and some way from your mouth, so it hears the room as much as it hears you. A headset or a USB microphone a hand's width from your mouth can make more difference than switching apps.

The room#

Background noise, music, other voices and echo from hard surfaces all force the model to guess. A furnished room is easier than a bare one, and a closed door beats an open-plan office.

Accents and languages#

Accuracy isn't the same for every speaker. OpenAI's Whisper paper found that how well the model does in a language tracks how much training audio it had for that language. The paper measures languages rather than accents, but the lesson carries over: don't assume an app that works well for a colleague will work as well for you. If English is your second language, or you dictate in a less widely spoken language, test with your own voice before trusting any general claim.

Words the model rarely sees#

Names, product names, acronyms and jargon often trip up general-purpose models. A model that has rarely heard "kubectl" or your colleague Siobhán will write something that sounds similar. A custom vocabulary (some apps call it a dictionary) tells the app which spellings to expect. Apple Dictation has no custom vocabulary; many dedicated dictation apps do.

How you speak#

Modern models use the words around an unclear word to work out what it was, so speaking in full phrases at a normal pace usually works better than slow, over-careful dictation. Keep the microphone at a steady distance, and pause between thoughts rather than in the middle of a sentence.

Cleanup after transcription#

Newer apps run a second step: a language model reads the transcript, removes filler words, applies the corrections you made mid-sentence, adds structure and fixes grammar. It doesn't change what the speech model heard, but it changes how much editing you're left with. It also adds a small risk: a cleanup model can occasionally change something you meant, so read important text before you send it.

What Published Research Says#

Independent, like-for-like comparisons of commercial dictation apps are rare, and vendors' own benchmarks can't be checked. The most useful public evidence is about the underlying models.

Whisper is robust, but not uniform. OpenAI's 2022 paper, Robust Speech Recognition via Large-Scale Weak Supervision, trained models on 680,000 hours of audio collected from the web. Its central finding is that a model trained on a large, varied mix of audio holds up better on unfamiliar recordings than models tuned to a single clean benchmark. It also shows that larger models do better, and that accuracy differs a lot from one language to another.

Errors aren't always small. The paper's limitations section describes failure modes that go beyond a misheard word: getting stuck repeating text, missing the first or last words of a segment, and occasionally producing text that isn't in the audio at all. They're a good reason to glance over anything important before you send it.

Weigh dictation errors against typing errors. A 2016 study by Stanford University, the University of Washington and Baidu found speech input about three times faster than typing on a smartphone keyboard, with fewer errors. It compared thumb typing with the speech recognition of the time, not today's desktop apps, but it's a reminder that the alternative to dictation isn't perfect text; it's your own typing, typos included.

Public leaderboards compare models, not apps. The Hugging Face Open ASR Leaderboard ranks open speech-recognition models, Whisper among them, by error rate and speed on public test sets, with English, multilingual and long-form tracks. It's a good way to see how model families compare. It doesn't test commercial dictation apps, their cleanup steps or your microphone.

Benchmarks are a starting point. Results on public test sets tell you which models are capable. They don't tell you how an app will cope with your microphone, your room and your vocabulary. That takes your own test.

How to Test Dictation Apps Yourself#

A fair comparison doesn't need special equipment. It needs a script, a quiet half hour and a little discipline about keeping conditions the same.

1. Write three short passages, each about 100–150 words:

• Everyday: an email or message you'd really send.
• Your vocabulary: the jargon, product names and acronyms you actually use.
• Names and numbers: people, places, dates, prices and phone numbers.

2. Keep the conditions the same. Same microphone, same distance, same room, same time of day. Read at your normal pace, as you would when writing for real.

3. Read each passage into each app, twice. Paste the results into a document next to your script.

4. Count two things for each result:

• Errors: substituted, missing and extra words compared with your script. Divide by the number of words in the script to get your own word error rate.
• Fixes to send: the edits you'd make before sending it. This is the number that decides whether dictation saves you time.

5. Retry the passage that went worst after adding your problem words to the app's custom vocabulary, if it has one.

6. Add one real-world condition: a noisier room, a longer passage (a minute or more of talking), or a second language if you use one.

A simple scoring sheet:

For apps that transcribe recorded files, such as MacWhisper, you can go one better: record each passage once and give the same audio file to every app, so the only thing that changes is the software.

If you're comfortable with Python, the open-source jiwer library calculates word error rate from your script and a transcript, so you don't have to count by hand.

How the Apps Differ#

Without trustworthy head-to-head numbers, the useful comparison is what each app documents about the things that drive accuracy.

Dragon, long the reference for legal and medical dictation, is no longer an option on a Mac: Dragon for Mac was discontinued in 2018, and Dragon Professional v16 runs on Windows only.

What the table means in practice:

• Live dictation into any app, with cleanup: LumeVoice and Wispr Flow are both cloud-based, with dictionaries and automatic cleanup. Both need an internet connection.
• Audio that must stay on your Mac: Superwhisper with a local model, or MacWhisper, keeps processing offline. The largest models need a capable Mac; MacWhisper's own page says the Medium and Large Whisper models need more than 8 GB of RAM to work well.
• Short messages at no cost: Apple Dictation is built in and can process speech on your Mac (Keyboard settings show which languages it handles on the device), but you'll fix names, jargon and filler words yourself.
• Legal and medical vocabulary: test hardest here. Add your terms to a custom vocabulary and check every output that matters; no app removes the need for review in those fields.

Limitations of Any Accuracy Comparison#

1. Models change. Apps update their speech models and cleanup steps, often without much notice, so any comparison is a snapshot. Retest after major updates.
2. Your voice isn't the test set. Results for one speaker, accent or microphone don't transfer reliably to another.
3. Verbatim and clean are different goals. A tool built for word-for-word transcripts (interviews, subtitles) and one built for ready-to-send writing should be judged differently.
4. Live dictation is hard to test fairly. You can't say anything exactly the same way twice. Reading a fixed script twice per app, or feeding the same recording to file-based tools, reduces the noise but doesn't remove it.

Sources#

• Radford et al., Robust Speech Recognition via Large-Scale Weak Supervision, OpenAI, 2022.
• Hugging Face, Open ASR Leaderboard.
• Ruan et al., 2016, Stanford University, University of Washington and Baidu: Stanford's report on speech input vs typing on smartphones.
• Features as documented on each company's site, checked October 2026: Wispr Flow, Superwhisper, MacWhisper.

Further Reading#

• The most accurate dictation app for Mac
• How to fix poor dictation accuracy on Mac
• Voice dictation vs typing: how much faster is speaking?
• Best AI dictation apps for Mac
• Best Wispr Flow alternatives
• MacWhisper review
• LumeVoice vs Superwhisper
• Android Voice Typing Not Working? 9 Fixes for Gboard and More
• Best Android Dictation Apps in 2026 (Gboard and Beyond)
• Best Gboard Voice Typing Alternatives in 2026
• Windows Voice Typing Not Working? 10 Fixes in the Right Order