Almost every dictation and transcription vendor publishes an accuracy figure, and almost none of them explain how it was measured. Sonix claims 99%. Happy Scribe claims 99% for human transcription. BigHand claims up to 99%. Whisper Memos claims 96.7%. Speechnotes claims around 95%. Windows Voice Typing sits around 85–90% on conversational English.
None of these numbers are audited, and they're not measured the same way. Here's what the metric actually is and how to test tools on the audio you care about.
Accuracy claims are the inverse of word error rate (WER), the standard metric in speech recognition:
WER = (substitutions + deletions + insertions) / total words in reference
Say the reference transcript is "the patient has a mild fever" (six words) and the system produces "the patient had a mild fever." That's one substitution: WER = 1/6 = 16.7%, so 83.3% accuracy.
Three things about this metric matter more than the number itself.
It weights every word equally. Getting "the" wrong counts the same as getting a drug name wrong. In clinical, legal, or technical work the errors that matter are concentrated in exactly the words a general model handles worst.
It ignores punctuation and casing in most implementations. A transcript with perfect words and no sentence breaks can score well and still be unreadable.
It can exceed 100%. A system that hallucinates a long passage generates more insertions than there are reference words. This isn't theoretical - Whisper-family models are known to invent fluent text during silence or noise.
| Claimed accuracy | WER | Errors per 1,000 words | Roughly |
|---|---|---|---|
| 99% | 1% | 10 | One error every two lines |
| 97% | 3% | 30 | One error per two sentences |
| 95% | 5% | 50 | One error per sentence |
| 90% | 10% | 100 | Two or three per sentence |
| 85% | 15% | 150 | Constant correction |
Note the top row. A 99% claim - the best figure anyone in this market advertises - still means roughly ten errors in a thousand words. That's a real editing pass, not a finished document.
Below about 95%, most people find correcting slower than typing would have been.
Published figures are measured on clean read speech: a single speaker, a good microphone, a quiet room, standard vocabulary, standard accent. Your audio is probably none of those things.
What degrades WER, in rough order of impact:
Microphone and distance. The largest single variable. A headset at six inches versus a laptop array across the desk can be the difference between 5% and 20% WER on the same voice, in the same room, with the same software.
Room acoustics. Reverb hurts more than steady background noise. Hard surfaces and a bare room are worse than a quiet fan.
Overlapping speech. Two people talking at once is where automated meeting transcription fails hardest, and where speaker diarization falls apart.
Accent and dialect. Models are trained on skewed data. WER on non-standard accents is routinely several times higher than on the accents the marketing figures were measured with. This is a well-documented equity problem, not a rare edge case.
Specialist vocabulary. Drug names, case citations, product names, and technical jargon are exactly the words a general model gets wrong - and exactly the words you can't afford to have wrong.
Audio quality. Compressed voice memos, phone recordings, and low-bitrate exports all cost you accuracy before the model sees anything.
Rev charges $1.99 per audio minute for human transcription against $0.25 for AI. Happy Scribe charges $2 a minute for human, with extra AI minutes at $0.20. Roughly an eight to ten times premium.
Human transcriptionists typically land at 98–99% on difficult audio where automated systems drop to 80% or worse - overlapping speech, strong accents, poor recordings, dense jargon. On clean single-speaker audio the gap narrows considerably, and paying the premium buys much less.
The decision rule that works: if two of {overlapping speakers, strong accents, bad recording, heavy jargon} apply, pay for a human on the sections that matter. Otherwise use AI and edit.
Traditional speech recognition failed by producing garbage - obviously wrong text you'd catch immediately. Whisper-family models fail differently: during silence, background noise, or unclear audio they sometimes generate fluent, plausible sentences that were never spoken.
This affects MacWhisper, Buzz, Vibe, superwhisper, and every other Whisper-based tool, along with cloud services built on the same family.
It matters because it's invisible. A garbled transcript announces itself; an invented sentence reads perfectly. For clinical, legal, or journalistic work this is the failure mode to watch for, and the reason to check any quote against the audio before publishing it.
Mitigations: trim silence before transcribing, use voice activity detection where the tool offers it, and be suspicious of fluent passages in sections you know were unclear.
Tools like Wispr Flow, Aqua Voice, and Amical run the transcript through a language model to remove filler words and tidy the prose. The output usually needs less editing.
But it's no longer a transcript. The model paraphrases, and paraphrase is not what you want when the words matter. For interviews you'll quote, court proceedings, or clinical records, use a tool that transcribes verbatim, or check the cleaned version against the original.
Vendor numbers are marketing. Twenty minutes gets you a figure that's actually about your work.
That last step is the point. Most people shopping for accuracy are choosing between tools when they should be buying a microphone - see the best microphone for dictation.
No independent audited WER benchmarks exist for this market, so the directory doesn't publish accuracy rankings - anything we listed would be a vendor claim repeated.
What can be said generally: large Whisper models, as used by superwhisper, MacWhisper, Handy, and Buzz, are at the top of the open field. Dragon Professional with a trained profile still leads on specialist vocabulary, because it adapts to your voice and domain in a way general models don't. Built-in options like Apple Dictation are competitive on clean audio and fall behind on accents and jargon. Windows Voice Typing is the weakest of the mainstream options at roughly 85–90%, with no custom vocabulary to fix it.
Below 5% (95% accuracy) is usable with editing. Below 2% is very good. Above 10%, correcting is usually slower than typing.
It's the best figure anyone advertises, and it still means about ten errors per thousand words - roughly one every two lines. It's a first draft, not a finished document.
On clean audio, large Whisper models and the major cloud engines are close. On specialist vocabulary with a trained profile, Dragon Professional still leads. Recording quality affects the result more than the choice between good tools.
Vendor figures are measured on clean read speech with a good microphone in a quiet room. Your microphone, room acoustics, accent, and vocabulary all move the number, usually downwards.
Whisper-family models can generate plausible text during silence or unclear audio. It's uncommon but hard to spot, because the invented text reads fluently. Check quotes against the audio before publishing.
Record three to five minutes of your real speech, transcribe it by hand as a reference, run it through each candidate tool, and count substitutions, deletions, and insertions against your reference.
Last verified: 10 August 2026.