Transcription Accuracy: What Affects It and How to Get the Best Result
"How accurate is it?" is the first question everyone asks about transcription, and the honest answer is: it depends. On clear, well-recorded speech, modern transcription is very good — typically 90 to 95% of words correct. On a noisy phone call with three people talking over each other, it can fall below 80%. The difference is not the software. It is the audio, the speakers, and the words themselves.
This guide explains what affects transcription accuracy, how accuracy is measured, and what you can do — before, during, and after recording — to get the best result from any transcription tool. You can transcribe audio with AudioToText and download the result with timestamps, then check it using the methods below.
How Accuracy Is Measured
Accuracy is usually reported as word error rate (WER): the number of wrong, missing, and extra words, divided by the number of words spoken. A WER of 5% means 95 words in 100 are right. That sounds good, and for ordinary listening it is — but the 5% is not spread evenly. It clusters on the words that matter most: names, numbers, technical terms, and short words like "not" that change the meaning of a sentence.
That clustering is why a single accuracy number never tells the whole story. A transcript that is 95% accurate can still be wrong about the one number in a meeting, the one name in an interview, or the one "not" in a contract discussion. Judging a transcript means checking the words that carry meaning, not counting the words that are right.
The National Institute of Standards and Technology has run speech recognition evaluations for decades, and its work is the reason the field measures accuracy the way it does. See NIST for the evaluation methods and the history behind them.
Why Accuracy Varies So Much
Six factors drive almost all of the difference between a good transcript and a bad one. They are worth knowing, because five of them are under your control.
1. Audio quality
This is the biggest factor by far. A clear recording made close to the microphone transcribes well. A distant, echoey, or noisy recording does not. Background music, air conditioning, traffic, and keyboard clatter all compete with the voice, and the transcription has to guess which sounds are speech. Every decibel of noise costs accuracy, and there is no software setting that recovers a voice buried in noise.
2. Overlapping speech
When two people talk at once, even the best systems struggle. The audio contains two voices at the same time, and the system must separate them, recognise both, and decide what order to write them in. It usually gets one of them, partly. Overlapping speech is the single largest source of errors in meeting transcripts, and it is also the most avoidable: one voice at a time is a habit, not a technical fix.
3. Accents and speaking style
Transcription is trained mostly on standard, clearly spoken English, and it is best at what it has heard most. Strong regional accents, fast speech, mumbling, and heavy use of filler words all lower accuracy. Speaking quickly is particularly costly: at 200 words a minute, words run together and the system has less time to separate them. Slowing down slightly — not dramatically, just to a normal conversational pace — helps more than any setting.
4. Vocabulary
Names, product names, technical terms, and jargon are the hardest words to transcribe, because the system has often never seen them. A rare surname, a drug name, a piece of engineering jargon — these are the words the system guesses at, and guesses wrong. Saying a name slowly, spelling it once, and using the full term instead of an abbreviation all help. This is also why checking the transcript matters: the errors cluster exactly where the vocabulary is unusual.
5. Recording setup
A phone held at arm's length, a laptop microphone across a room, a speakerphone in a conference room — these all produce the same problem: the voice is quiet relative to the room. A microphone close to the speaker fixes it. The difference between a built-in laptop microphone at two metres and a headset microphone at ten centimetres is often ten points of accuracy or more.
6. The language and the model
Accuracy is best in English and other widely spoken languages, where the training data is largest. Less common languages, and meetings that switch between languages, are harder. Larger transcription models are generally more accurate than smaller ones, but they are also slower — which is why most services use a mid-size model by default and reserve the largest for the hardest audio.
How to Get the Best Transcript: Before You Record
| Step | What to do | Why it helps |
|---|---|---|
| Microphone | Use a headset or a USB mic, close to the speaker | The voice is loud relative to the room |
| Room | Choose a quiet room, turn off anything that hums | Less noise competing with the voice |
| Test | Record 30 seconds and listen back | Catches a bad setup before the real recording |
| Introductions | Say each person's name at the start, spelled out | Fixes the most common error before it happens |
| Numbers | Say full dates and amounts, not shorthand | "March 15" transcribes better than "the fifteenth" |
| One voice | Agree to speak one at a time | Removes the largest source of errors |
These six steps take five minutes and are worth more than any amount of editing afterwards. A clean recording is the only reliable way to a clean transcript.
How to Get the Best Transcript: While You Record
Speak at a normal pace, not a rushed one. Finish one sentence before starting the next. Pause briefly between topics — the pause gives the transcription a clean boundary, and it gives the listener a moment to catch up. If a noise covers a line — a door slamming, a phone ringing — repeat the line once the noise has passed. Say it out loud: "Let me repeat that, the door covered it." The repetition is the most valuable line in the transcript, because it is the line that would otherwise have been lost.
If someone says something important, restate it slowly: "So what we agreed is…" That sentence is easy to transcribe, easy to find in the text, and easy to quote later. It also confirms the point while everyone is still in the room to correct it.
How to Check a Transcript After Transcription
Checking is where accuracy is actually decided. A transcript you do not check is a transcript you are guessing about. The check does not take long if you know what to look for.
Step 1: Read it once, fast
Skim the whole transcript. Mark anything that looks wrong: a sentence that makes no sense, a word that does not fit, a passage where the text jumps. Most errors are visible on a fast read, because wrong words break the sense of a sentence.
Step 2: Check the words that carry meaning
Go through the names, the numbers, the dates, and the words "not", "no", and "never". Use the timestamps to jump to the audio and listen to those exact passages. A name spelled wrong is embarrassing; a number wrong is misleading; a missing "not" reverses the meaning. This check takes ten minutes for an hour of audio, and it is the most valuable ten minutes in the process.
Step 3: Listen to the unclear passages
Where the transcript marks a passage as unclear, or where the text is garbled, listen to the audio at that timestamp. If you can hear the words, fix the text. If you cannot, mark the passage as inaudible with the timestamp, and move on. Never guess a word you cannot hear — a guessed word presented as fact is worse than an honest gap.
Step 4: Compare the length
A quick sanity check: count the words in the transcript and compare them with the recording length. Spoken English runs at roughly 130 to 160 words a minute. An hour of speech should give you roughly 8,000 to 9,500 words. If your transcript has far fewer, something is missing — a section where the audio was too quiet, or a speaker the system did not hear. If it has far more, the system may have repeated itself, and the transcript needs a closer look.
What the Numbers Mean in Practice
| Audio condition | Typical accuracy | What it feels like |
|---|---|---|
| Studio, one speaker, close mic | 95–98% | A word wrong here and there; names may need checking |
| Clear meeting, one voice at a time | 90–95% | Readable; check names and numbers |
| Ordinary phone call | 85–90% | Readable with effort; some sentences need the audio |
| Noisy room, distant mic | 75–85% | Rough; use it as a guide, not a record |
| Heavy overlap, strong noise | Below 75% | Fragments; re-record if the content matters |
These are typical ranges, not promises. Your recording will land somewhere in the range that matches its conditions, and the checking steps above are what move it from the range to a usable transcript.
Common Mistakes That Cost Accuracy
Recording with the laptop microphone across the table. The single most common mistake, and the easiest to fix. Move the microphone closer, or use a headset.
Letting people talk over each other. In meetings, this is a habit, not an accident. A one-line reminder at the start — "One voice at a time, please" — fixes most of it.
Trusting the transcript without checking it. A transcript is a draft until it is checked. The words that matter are the words most likely to be wrong, so check those first, every time.
Recording in a noisy place when a quiet one is available. A café is a bad recording studio. Step outside, into a corridor, or into a parked car. Thirty seconds of quiet is worth an hour of editing.
Using the transcript for something it cannot do. A transcript of a noisy, overlapping meeting is not a legal record and not a publishable quote. If the content matters that much, re-record it properly — or accept that the transcript is a guide, and listen to the audio for the passages that count. For guidance on when a transcript is good enough to publish, read our guide to transcribing interviews.
Frequently Asked Questions
Is 95% accuracy good enough?
For most uses, yes — with checking. A 95% transcript is very readable, and the errors are mostly on words that do not change the meaning. For publication, for legal use, or for anything where a single word matters, check the transcript against the audio and fix the words that carry meaning.
Why does transcription get names wrong?
Because names are rare words the system has often never seen. It guesses the closest familiar word instead. Saying a name slowly and spelling it once at the start of the recording is the best fix, followed by checking names in the transcript.
Does a bigger model mean better accuracy?
Usually, yes — larger models are more accurate, especially on difficult audio. They are also slower and more expensive to run. For clear audio, the difference between a mid-size model and the largest one is small. For noisy or accented audio, it is larger. If your audio is hard, it is worth using the best model you can; if it is clean, the default is fine.
Can accuracy be improved after transcription?
Partly. Checking and fixing the transcript helps, and a second pass with the audio playing catches errors a silent read misses. But no amount of editing can recover words that were never captured because the audio was too noisy. Accuracy is decided at recording time; editing only polishes it.
What about transcription in other languages?
Accuracy is best in English and other widely spoken languages. For less common languages, expect more errors, and check the transcript more carefully. If a recording switches languages, expect the errors to cluster at the switching points. See pricing for the plans — the Free plan covers 60 minutes a month, enough to test accuracy on your own recordings before you commit.
Key Takeaways
- Accuracy is decided at recording time: a close microphone, a quiet room, and one voice at a time matter more than any setting.
- Check the words that carry meaning — names, numbers, dates, and "not" — against the audio, using the timestamps.
- Never guess a word you cannot hear; mark it inaudible and move on.
- A 90–95% transcript is very usable, but it is a draft until it is checked.
- If the content matters and the recording is bad, re-record it. No editing recovers words that were never captured.
!Transcription accuracy factors
Image description: The image should show a simple bar chart with five bars labeled "Clear audio", "Close microphone", "One voice at a time", "Slow clear speech", and "Familiar vocabulary". The bars for clear audio and close microphone are the tallest, showing they matter most. A caption reads "Five things decide accuracy — and you control all five."
References
- National Institute of Standards and Technology — speech recognition evaluation: https://www.nist.gov/
- National Institute on Deafness and Other Communication Disorders — hearing statistics: https://www.nidcd.nih.gov/
- Library of Congress — audio collections and transcripts: https://www.loc.gov/