Every dictation product on the market quotes an accuracy figure, and almost none of them explain what it means. Ninety-five per cent sounds close to perfect. It also sounds a lot like ninety per cent, which is a completely different daily experience. So let us be specific about what these numbers feel like once you are actually using them.
The short version: on clear speech in a normal room, good dictation software lands around 90 to 95 per cent. That is roughly one error every twenty words, and most errors are trivial. The gap between a frustrating experience and a brilliant one is rarely the software. It is nearly always the microphone, the language setting and whether you have taught it your own vocabulary.
What the percentages actually mean
Accuracy is measured per word. At 95 per cent, one word in twenty comes out wrong. A typical spoken sentence runs to about fifteen words, so you are looking at an error every sentence or two. At 90 per cent it is one word in ten, which is closer to one or two per sentence.
Those sound alarming written down. In practice they are far less painful than they look, for one simple reason: most errors are obvious. When "Kowalczyk" comes out as three odd syllables, you spot it instantly. The dangerous errors are the plausible ones, where a real word replaces another real word and the sentence still reads sensibly. Those are rarer, but they are the reason you should always read back anything important.
Here is the comparison nobody makes, though. Human typing is not 100 per cent either. Studies of ordinary typists put raw error rates somewhere between two and eight per cent before correction. Dictation at 95 per cent is not competing with perfection. It is competing with your actual typing, at roughly three times the speed.
The five things that quietly wreck accuracy
1. The microphone in your laptop lid
This is the big one, and it is boring, which is why people ignore it. A built-in laptop microphone sits centimetres from a whirring fan, points at the ceiling rather than your mouth, and picks up the whole room. A twenty pound headset will do more for your accuracy than any software setting, any model size, and any amount of speaking carefully. If you change one thing after reading this, change this.
2. The wrong language setting
A model expecting English will do something strange and confident with French. It is a surprisingly common cause of "this software is rubbish", and it takes two seconds to fix. If you work across languages regularly, our guide to multilingual voice typing covers how to switch cleanly.
3. Speaking too carefully
This one is counter-intuitive, so it catches nearly everybody. People assume that slowing down and enunciating each word will help. It does the opposite. These models are trained on natural conversational speech, and they use surrounding words as context to decide what they heard. Strip out the natural rhythm and you strip out the context. Talk as if you were explaining something to a colleague and you will get better results than if you talk like a robot.
4. Words the model has barely heard
Generic speech recognition is trained on generic speech. Your colleague's surname, your industry's acronyms, the drug name you write fifteen times a day, the product your company sells: these are, almost by definition, the words it has heard least. This is exactly what custom vocabulary is for, and adding twenty terms takes five minutes and pays for itself the same afternoon.
5. Rooms with things going on in them
Background chatter, a television, an open-plan office, a fan heater. Noise does not just add sound, it masks the quieter consonants that distinguish similar words. A headset with a directional microphone helps enormously here, because it hears you and not the room.
What pushes accuracy up
Beyond fixing the five problems above, two levers are worth knowing about.
Model size. PeekoType ships five, from Tiny to Large. Bigger models are measurably more accurate, particularly with accents, technical terms and imperfect audio, and the cost is speed. On a modern machine, moving from Base to Small is often free in practical terms and noticeably better. Our piece on running offline AI on older PCs goes into how to choose.
Punctuation and cleanup. A transcript is not just words, it is readable text. Automatic punctuation, filler word removal and capitalisation do not change the raw recognition rate, but they change how accurate the result feels, which is what you actually care about.
Accents, and the question everyone asks
Yes, regional accents work. Modern models were trained on enormous volumes of varied speech rather than on a single reference accent, so Glaswegian, Geordie, Scouse, Yorkshire, West Country, Welsh and Northern Irish speakers generally get on fine. Accuracy is usually a shade lower than for a neutral southern English accent, but the gap is much smaller than it was five years ago, and a good microphone closes most of it.
Second-language English speakers also do well, which surprises people. The model has heard a great deal of accented English and is not thrown by it.
A fair test you can run in ten minutes
If you want to know how a dictation tool performs for you, rather than for a marketing page, do this.
- Put on a headset. Not the laptop microphone. This is the point of the test.
- Set the language correctly.
- Dictate roughly two hundred words of something you would genuinely write, not a scripted passage. Real work has real vocabulary in it.
- Count the errors. Divide by the word count. That is your actual accuracy.
- Now add your ten most-used specialist terms to the vocabulary list and dictate the same thing again.
Most people find the second run is dramatically better, and that the remaining errors cluster around a handful of predictable words they can simply add to the list.
The honest verdict
Voice typing is not flawless and anyone claiming otherwise is selling something. What it is, reliably, is faster than typing for anything longer than a sentence, accurate enough that corrections are quick, and genuinely transformative if you have sore wrists, dyslexia, or simply more to write than hours to write it in.
PeekoType runs at around 90 to 95 per cent on clear speech, entirely on your own computer, with five model sizes and a vocabulary list you control. There is a free 14-day trial with no card required, which is the only accuracy test that means anything.